billion-context 0.1.87 → 0.1.88-pr.571.415

package/dist/index.js

@@ -9365,7 +9365,7 @@ var require_pool = __commonJS({
     function defaultFactory(origin, opts) {
       return new Client(origin, opts);
     }
-    var Pool = class extends PoolBase {
+    var Pool2 = class extends PoolBase {
       constructor(origin, {
         connections,
         factory = defaultFactory,
@@ -9438,7 +9438,7 @@ var require_pool = __commonJS({
         }
       }
     };
-    module.exports = Pool;
+    module.exports = Pool2;
   }
 });
 
@@ -9458,7 +9458,7 @@ var require_balanced_pool = __commonJS({
       kRemoveClient,
       kGetDispatcher
     } = require_pool_base();
-    var Pool = require_pool();
+    var Pool2 = require_pool();
     var { kUrl } = require_symbols();
     var util = require_util();
     var kFactory = /* @__PURE__ */ Symbol("factory");
@@ -9479,7 +9479,7 @@ var require_balanced_pool = __commonJS({
       return a;
     }
     function defaultFactory(origin, opts) {
-      return new Pool(origin, opts);
+      return new Pool2(origin, opts);
     }
     var BalancedPool = class extends PoolBase {
       constructor(upstreams = [], { factory = defaultFactory, ...opts } = {}) {
@@ -9714,7 +9714,7 @@ var require_agent = __commonJS({
     var { InvalidArgumentError, MaxOriginsReachedError } = require_errors();
     var { kClients, kRunning, kClose, kDestroy, kDispatch, kUrl } = require_symbols();
     var DispatcherBase = require_dispatcher_base();
-    var Pool = require_pool();
+    var Pool2 = require_pool();
     var Client = require_client();
     var util = require_util();
     var kOnConnect = /* @__PURE__ */ Symbol("onConnect");
@@ -9725,9 +9725,9 @@ var require_agent = __commonJS({
     var kOptions = /* @__PURE__ */ Symbol("options");
     var kOrigins = /* @__PURE__ */ Symbol("origins");
     function defaultFactory(origin, opts) {
-      return opts && opts.connections === 1 ? new Client(origin, opts) : new Pool(origin, opts);
+      return opts && opts.connections === 1 ? new Client(origin, opts) : new Pool2(origin, opts);
     }
-    var Agent = class extends DispatcherBase {
+    var Agent2 = class extends DispatcherBase {
       constructor({ factory = defaultFactory, maxOrigins = Infinity, connect, ...options } = {}) {
         if (typeof factory !== "function") {
           throw new InvalidArgumentError("factory must be a function.");
@@ -9836,7 +9836,7 @@ var require_agent = __commonJS({
         return allClientStats;
       }
     };
-    module.exports = Agent;
+    module.exports = Agent2;
   }
 });
 
@@ -10344,7 +10344,7 @@ var require_socks5_proxy_agent = __commonJS({
     var { InvalidArgumentError } = require_errors();
     var { Socks5Client, STATES } = require_socks5_client();
     var { kDispatch, kClose, kDestroy } = require_symbols();
-    var Pool = require_pool();
+    var Pool2 = require_pool();
     var buildConnector = require_connect();
     var { debuglog } = __require("util");
     var debug = debuglog("undici:socks5-proxy");
@@ -10467,7 +10467,7 @@ var require_socks5_proxy_agent = __commonJS({
           const originKey = String(origin);
           let pool = this[kPools].get(originKey);
           if (!pool || pool.destroyed || pool.closed) {
-            pool = new Pool(origin, {
+            pool = new Pool2(origin, {
               pipelining: opts.pipelining,
               connections: opts.connections,
               connect: async (connectOpts, callback) => {
@@ -10542,8 +10542,8 @@ var require_proxy_agent = __commonJS({
   "node_modules/undici/lib/dispatcher/proxy-agent.js"(exports, module) {
     "use strict";
     var { kProxy, kClose, kDestroy, kDispatch } = require_symbols();
-    var Agent = require_agent();
-    var Pool = require_pool();
+    var Agent2 = require_agent();
+    var Pool2 = require_pool();
     var DispatcherBase = require_dispatcher_base();
     var { InvalidArgumentError, RequestAbortedError, SecureProxyConnectionError } = require_errors();
     var buildConnector = require_connect();
@@ -10561,7 +10561,7 @@ var require_proxy_agent = __commonJS({
       return protocol === "https:" ? 443 : 80;
     }
     function defaultFactory(origin, opts) {
-      return new Pool(origin, opts);
+      return new Pool2(origin, opts);
     }
     var noop = () => {
     };
@@ -10569,7 +10569,7 @@ var require_proxy_agent = __commonJS({
       if (opts.connections === 1) {
         return new Client(origin, opts);
       }
-      return new Pool(origin, opts);
+      return new Pool2(origin, opts);
     }
     var Http1ProxyWrapper = class extends DispatcherBase {
       #client;
@@ -10673,7 +10673,7 @@ var require_proxy_agent = __commonJS({
         } else {
           this[kClient] = clientFactory(url, { connect });
         }
-        this[kAgent] = new Agent({
+        this[kAgent] = new Agent2({
           ...opts,
           factory,
           connect: async (opts2, callback) => {
@@ -10801,7 +10801,7 @@ var require_env_http_proxy_agent = __commonJS({
     var DispatcherBase = require_dispatcher_base();
     var { kClose, kDestroy, kClosed, kDestroyed, kDispatch, kNoProxyAgent, kHttpProxyAgent, kHttpsProxyAgent } = require_symbols();
     var ProxyAgent2 = require_proxy_agent();
-    var Agent = require_agent();
+    var Agent2 = require_agent();
     var DEFAULT_PORTS = {
       "http:": 80,
       "https:": 443
@@ -10814,7 +10814,7 @@ var require_env_http_proxy_agent = __commonJS({
         super();
         this.#opts = opts;
         const { httpProxy, httpsProxy, noProxy, ...agentOpts } = opts;
-        this[kNoProxyAgent] = new Agent(agentOpts);
+        this[kNoProxyAgent] = new Agent2(agentOpts);
         const HTTP_PROXY = httpProxy ?? process.env.http_proxy ?? process.env.HTTP_PROXY;
         if (HTTP_PROXY) {
           this[kHttpProxyAgent] = new ProxyAgent2({ ...agentOpts, uri: HTTP_PROXY });
@@ -13384,7 +13384,7 @@ var require_mock_pool = __commonJS({
   "node_modules/undici/lib/mock/mock-pool.js"(exports, module) {
     "use strict";
     var { promisify: promisify2 } = __require("util");
-    var Pool = require_pool();
+    var Pool2 = require_pool();
     var { buildMockDispatch } = require_mock_utils();
     var {
       kDispatches,
@@ -13399,7 +13399,7 @@ var require_mock_pool = __commonJS({
     var { MockInterceptor } = require_mock_interceptor();
     var Symbols = require_symbols();
     var { InvalidArgumentError } = require_errors();
-    var MockPool = class extends Pool {
+    var MockPool = class extends Pool2 {
       constructor(origin, opts) {
         if (!opts || !opts.agent || typeof opts.agent.dispatch !== "function") {
           throw new InvalidArgumentError("Argument opts.agent must implement Agent");
@@ -13486,7 +13486,7 @@ var require_mock_agent = __commonJS({
   "node_modules/undici/lib/mock/mock-agent.js"(exports, module) {
     "use strict";
     var { kClients } = require_symbols();
-    var Agent = require_agent();
+    var Agent2 = require_agent();
     var {
       kAgent,
       kMockAgentSet,
@@ -13524,7 +13524,7 @@ var require_mock_agent = __commonJS({
         if (opts?.agent && typeof opts.agent.dispatch !== "function") {
           throw new InvalidArgumentError("Argument opts.agent must implement Agent");
         }
-        const agent = opts?.agent ? opts.agent : new Agent(opts);
+        const agent = opts?.agent ? opts.agent : new Agent2(opts);
         this[kAgent] = agent;
         this[kClients] = agent[kClients];
         this[kOptions] = mockOptions;
@@ -14131,7 +14131,7 @@ var require_snapshot_recorder = __commonJS({
 var require_snapshot_agent = __commonJS({
   "node_modules/undici/lib/mock/snapshot-agent.js"(exports, module) {
     "use strict";
-    var Agent = require_agent();
+    var Agent2 = require_agent();
     var MockAgent = require_mock_agent();
     var { SnapshotRecorder } = require_snapshot_recorder();
     var WrapHandler = require_wrap_handler();
@@ -14182,7 +14182,7 @@ var require_snapshot_agent = __commonJS({
         });
         this[kSnapshotLoaded] = false;
         if (this[kSnapshotMode] === "record" || this[kSnapshotMode] === "update" || this[kSnapshotMode] === "playback" && opts.excludeUrls && opts.excludeUrls.length > 0) {
-          this[kRealAgent] = new Agent(opts);
+          this[kRealAgent] = new Agent2(opts);
         }
         if ((this[kSnapshotMode] === "playback" || this[kSnapshotMode] === "update") && this[kSnapshotPath]) {
           this.loadSnapshots().catch(() => {
@@ -14421,9 +14421,9 @@ var require_global2 = __commonJS({
     var globalDispatcher = /* @__PURE__ */ Symbol.for("undici.globalDispatcher.2");
     var legacyGlobalDispatcher = /* @__PURE__ */ Symbol.for("undici.globalDispatcher.1");
     var { InvalidArgumentError } = require_errors();
-    var Agent = require_agent();
+    var Agent2 = require_agent();
     if (getGlobalDispatcher() === void 0) {
-      setGlobalDispatcher(new Agent());
+      setGlobalDispatcher(new Agent2());
     }
     function setGlobalDispatcher(agent) {
       if (!agent || typeof agent.dispatch !== "function") {
@@ -18828,7 +18828,7 @@ var require_headers = __commonJS({
         }
       }
     };
-    var Headers = class _Headers {
+    var Headers2 = class _Headers {
       #guard;
       /**
        * @type {HeadersList}
@@ -18969,13 +18969,13 @@ var require_headers = __commonJS({
         target.#headersList = list;
       }
     };
-    var { getHeadersGuard, setHeadersGuard, getHeadersList, setHeadersList } = Headers;
-    Reflect.deleteProperty(Headers, "getHeadersGuard");
-    Reflect.deleteProperty(Headers, "setHeadersGuard");
-    Reflect.deleteProperty(Headers, "getHeadersList");
-    Reflect.deleteProperty(Headers, "setHeadersList");
-    iteratorMixin("Headers", Headers, headersListSortAndCombine, 0, 1);
-    Object.defineProperties(Headers.prototype, {
+    var { getHeadersGuard, setHeadersGuard, getHeadersList, setHeadersList } = Headers2;
+    Reflect.deleteProperty(Headers2, "getHeadersGuard");
+    Reflect.deleteProperty(Headers2, "setHeadersGuard");
+    Reflect.deleteProperty(Headers2, "getHeadersList");
+    Reflect.deleteProperty(Headers2, "setHeadersList");
+    iteratorMixin("Headers", Headers2, headersListSortAndCombine, 0, 1);
+    Object.defineProperties(Headers2.prototype, {
       append: kEnumerableProperty,
       delete: kEnumerableProperty,
       get: kEnumerableProperty,
@@ -18993,7 +18993,7 @@ var require_headers = __commonJS({
     webidl.converters.HeadersInit = function(V2, prefix, argument) {
       if (webidl.util.Type(V2) === webidl.util.Types.OBJECT) {
         const iterator = Reflect.get(V2, Symbol.iterator);
-        if (!util.types.isProxy(V2) && iterator === Headers.prototype.entries) {
+        if (!util.types.isProxy(V2) && iterator === Headers2.prototype.entries) {
           try {
             return getHeadersList(V2).entriesList;
           } catch {
@@ -19014,7 +19014,7 @@ var require_headers = __commonJS({
       fill: fill2,
       // for test.
       compareHeaderName,
-      Headers,
+      Headers: Headers2,
       HeadersList,
       getHeadersGuard,
       setHeadersGuard,
@@ -19028,7 +19028,7 @@ var require_headers = __commonJS({
 var require_response = __commonJS({
   "node_modules/undici/lib/web/fetch/response.js"(exports, module) {
     "use strict";
-    var { Headers, HeadersList, fill: fill2, getHeadersGuard, setHeadersGuard, setHeadersList } = require_headers();
+    var { Headers: Headers2, HeadersList, fill: fill2, getHeadersGuard, setHeadersGuard, setHeadersList } = require_headers();
     var { extractBody, cloneBody, mixinBody, streamRegistry, bodyUnusable } = require_body();
     var util = require_util();
     var nodeUtil = __require("util");
@@ -19104,7 +19104,7 @@ var require_response = __commonJS({
         }
         init = webidl.converters.ResponseInit(init);
         this.#state = makeResponse({});
-        this.#headers = new Headers(kConstruct);
+        this.#headers = new Headers2(kConstruct);
         setHeadersGuard(this.#headers, "response");
         setHeadersList(this.#headers, this.#state.headersList);
         let bodyWithType = null;
@@ -19379,7 +19379,7 @@ var require_response = __commonJS({
     function fromInnerResponse(innerResponse, guard) {
       const response = new Response2(kConstruct);
       setResponseState(response, innerResponse);
-      const headers = new Headers(kConstruct);
+      const headers = new Headers2(kConstruct);
       setResponseHeaders(response, headers);
       setHeadersList(headers, innerResponse.headersList);
       setHeadersGuard(headers, guard);
@@ -19451,7 +19451,7 @@ var require_request2 = __commonJS({
   "node_modules/undici/lib/web/fetch/request.js"(exports, module) {
     "use strict";
     var { extractBody, mixinBody, cloneBody, bodyUnusable } = require_body();
-    var { Headers, fill: fillHeaders, HeadersList, setHeadersGuard, getHeadersGuard, setHeadersList, getHeadersList } = require_headers();
+    var { Headers: Headers2, fill: fillHeaders, HeadersList, setHeadersGuard, getHeadersGuard, setHeadersList, getHeadersList } = require_headers();
     var util = require_util();
     var nodeUtil = __require("util");
     var {
@@ -19720,7 +19720,7 @@ var require_request2 = __commonJS({
             requestFinalizer.register(ac, { signal, abort }, abort);
           }
         }
-        this.#headers = new Headers(kConstruct);
+        this.#headers = new Headers2(kConstruct);
         setHeadersList(this.#headers, request.headersList);
         setHeadersGuard(this.#headers, "request");
         if (mode === "no-cors") {
@@ -20061,7 +20061,7 @@ var require_request2 = __commonJS({
       setRequestState(request, innerRequest);
       setRequestDispatcher(request, dispatcher);
       setRequestSignal(request, signal);
-      const headers = new Headers(kConstruct);
+      const headers = new Headers2(kConstruct);
       setRequestHeaders(request, headers);
       setHeadersList(headers, innerRequest.headersList);
       setHeadersGuard(headers, guard);
@@ -22538,8 +22538,8 @@ var require_cookies = __commonJS({
     var { parseSetCookie } = require_parse();
     var { stringify } = require_util4();
     var { webidl } = require_webidl();
-    var { Headers } = require_headers();
-    var brandChecks = webidl.brandCheckMultiple([Headers, globalThis.Headers].filter(Boolean));
+    var { Headers: Headers2 } = require_headers();
+    var brandChecks = webidl.brandCheckMultiple([Headers2, globalThis.Headers].filter(Boolean));
     function getCookies(headers) {
       webidl.argumentLengthCheck(arguments, 1, "getCookies");
       brandChecks(headers);
@@ -23280,7 +23280,7 @@ var require_connection = __commonJS({
     var { parseExtensions, isClosed, isClosing, isEstablished, isConnecting, validateCloseCodeAndReason } = require_util5();
     var { makeRequest } = require_request2();
     var { fetching } = require_fetch();
-    var { Headers, getHeadersList } = require_headers();
+    var { Headers: Headers2, getHeadersList } = require_headers();
     var { getDecodeSplit } = require_util2();
     var { WebsocketFrameSend } = require_frame();
     var assert = __require("assert");
@@ -23302,7 +23302,7 @@ var require_connection = __commonJS({
         useURLCredentials: true
       });
       if (options.headers) {
-        const headersList = getHeadersList(new Headers(options.headers));
+        const headersList = getHeadersList(new Headers2(options.headers));
         request.headersList = headersList;
       }
       const keyValue = crypto2.randomBytes(16).toString("base64");
@@ -25411,10 +25411,10 @@ var require_undici = __commonJS({
     "use strict";
     var Client = require_client();
     var Dispatcher = require_dispatcher();
-    var Pool = require_pool();
+    var Pool2 = require_pool();
     var BalancedPool = require_balanced_pool();
     var RoundRobinPool = require_round_robin_pool();
-    var Agent = require_agent();
+    var Agent2 = require_agent();
     var ProxyAgent2 = require_proxy_agent();
     var Socks5ProxyAgent = require_socks5_proxy_agent();
     var EnvHttpProxyAgent = require_env_http_proxy_agent();
@@ -25438,10 +25438,10 @@ var require_undici = __commonJS({
     Object.assign(Dispatcher.prototype, api);
     module.exports.Dispatcher = Dispatcher;
     module.exports.Client = Client;
-    module.exports.Pool = Pool;
+    module.exports.Pool = Pool2;
     module.exports.BalancedPool = BalancedPool;
     module.exports.RoundRobinPool = RoundRobinPool;
-    module.exports.Agent = Agent;
+    module.exports.Agent = Agent2;
     module.exports.ProxyAgent = ProxyAgent2;
     module.exports.Socks5ProxyAgent = Socks5ProxyAgent;
     module.exports.EnvHttpProxyAgent = EnvHttpProxyAgent;
@@ -47245,7 +47245,7 @@ function closeLogger() {
 }
 
 // src/upstream-proxy.ts
-var import_undici = __toESM(require_undici(), 1);
+var import_undici2 = __toESM(require_undici(), 1);
 import { execFileSync } from "child_process";
 import net from "net";
 import tls from "tls";
@@ -47323,6 +47323,179 @@ function maskHostInText(text, host) {
   return text;
 }
 
+// src/fetch-util.ts
+var import_undici = __toESM(require_undici(), 1);
+var MAX_REQUEST_BYTES = 100 * 1024 * 1024;
+var UPSTREAM_TIMEOUT_MS = 12 * 60 * 1e3;
+var liveUpstreamTimers = /* @__PURE__ */ new Set();
+function upstreamTimeoutMs() {
+  const raw = Number(process.env.BILI_UPSTREAM_TIMEOUT_MS);
+  return Number.isInteger(raw) && raw > 0 ? raw : UPSTREAM_TIMEOUT_MS;
+}
+var directDispatchers = /* @__PURE__ */ new Map();
+function directDispatcher(timeoutMs) {
+  let agent = directDispatchers.get(timeoutMs);
+  if (!agent) {
+    agent = new import_undici.Agent({ headersTimeout: timeoutMs, bodyTimeout: timeoutMs });
+    directDispatchers.set(timeoutMs, agent);
+  }
+  return agent;
+}
+async function fetchWithTimeout(url, opts, timeoutMs, externalSignal) {
+  const effective = timeoutMs ?? upstreamTimeoutMs();
+  const controller = new AbortController();
+  let cleared = false;
+  const armTimer = () => {
+    const t = setTimeout(() => {
+      liveUpstreamTimers.delete(t);
+      controller.abort();
+    }, effective);
+    liveUpstreamTimers.add(t);
+    return t;
+  };
+  let timer3 = armTimer();
+  const rearm = () => {
+    if (cleared) return;
+    clearTimeout(timer3);
+    liveUpstreamTimers.delete(timer3);
+    timer3 = armTimer();
+  };
+  let onExternalAbort = null;
+  if (externalSignal) {
+    if (externalSignal.aborted) controller.abort();
+    else {
+      onExternalAbort = () => controller.abort();
+      externalSignal.addEventListener("abort", onExternalAbort, { once: true });
+    }
+  }
+  const cleanup = () => {
+    cleared = true;
+    clearTimeout(timer3);
+    liveUpstreamTimers.delete(timer3);
+    if (onExternalAbort && externalSignal) externalSignal.removeEventListener("abort", onExternalAbort);
+  };
+  try {
+    const finalOpts = {
+      ...opts,
+      signal: controller.signal,
+      dispatcher: opts.dispatcher ?? directDispatcher(effective)
+    };
+    const raw = await fetch(url, finalOpts);
+    if (raw.body) {
+      const wrapped = armIdleBody(raw.body, rearm);
+      return {
+        response: new Response(wrapped, {
+          status: raw.status,
+          statusText: raw.statusText,
+          headers: raw.headers
+        }),
+        clearTimer: cleanup
+      };
+    }
+    return { response: raw, clearTimer: cleanup };
+  } catch (e) {
+    cleanup();
+    throw e;
+  }
+}
+function armIdleBody(body, rearm) {
+  const reader = body.getReader();
+  return new ReadableStream({
+    async pull(controller) {
+      try {
+        const result = await reader.read();
+        if (result.done) {
+          controller.close();
+          return;
+        }
+        rearm();
+        controller.enqueue(result.value);
+      } catch (e) {
+        controller.error(e);
+      }
+    },
+    async cancel(reason) {
+      try {
+        await reader.cancel(reason);
+      } catch {
+      }
+    }
+  });
+}
+var UpstreamHttpError = class extends Error {
+  status;
+  body;
+  attempts;
+  constructor(status, body, attempts) {
+    super(`upstream error ${status}`);
+    this.name = "UpstreamHttpError";
+    this.status = status;
+    this.body = body;
+    this.attempts = attempts;
+  }
+};
+var TRANSIENT_BODY_MARKERS = [
+  "captcha",
+  "verify failed",
+  "risk control",
+  "\u98CE\u63A7",
+  "rate limit",
+  "too many requests",
+  "try again"
+];
+function isTransientUpstreamError(status, body) {
+  if (status === 429 || status >= 500) return true;
+  if (status < 400) return false;
+  const lower = body.toLowerCase();
+  return TRANSIENT_BODY_MARKERS.some((marker) => lower.includes(marker));
+}
+var REPLAY_MAX_ATTEMPTS = 3;
+function replayMaxAttempts() {
+  const raw = Number(process.env.BILI_REPLAY_RETRY_MAX);
+  return Number.isInteger(raw) && raw >= 1 ? raw : REPLAY_MAX_ATTEMPTS;
+}
+function replayBaseDelayMs() {
+  const raw = Number(process.env.BILI_REPLAY_RETRY_BASE_MS);
+  return Number.isFinite(raw) && raw >= 0 ? raw : 1500;
+}
+function maxShrinkPerCompress() {
+  const raw = Number(process.env.BILI_MAX_SHRINK_PER_COMPRESS);
+  return Number.isFinite(raw) && raw > 0 && raw <= 1 ? raw : void 0;
+}
+function replayBackoffMs(attempt) {
+  return replayBaseDelayMs() * 2 ** (attempt - 1);
+}
+function sleep(ms2, signal) {
+  if (ms2 <= 0 || signal?.aborted) return Promise.resolve();
+  return new Promise((resolve) => {
+    let timer3 = null;
+    const finish2 = () => {
+      if (timer3) clearTimeout(timer3);
+      if (signal) signal.removeEventListener("abort", finish2);
+      resolve();
+    };
+    timer3 = setTimeout(finish2, ms2);
+    if (signal) signal.addEventListener("abort", finish2, { once: true });
+  });
+}
+async function fetchWithRetry(url, opts, timeoutMs, externalSignal, onRetry) {
+  const maxAttempts = replayMaxAttempts();
+  for (let attempt = 1; ; attempt++) {
+    const result = await fetchWithTimeout(url, opts, timeoutMs, externalSignal);
+    if (result.response.ok) return result;
+    const errText2 = await result.response.text().catch(() => "upstream error");
+    result.clearTimer();
+    const lastAttempt = attempt >= maxAttempts;
+    if (!lastAttempt && isTransientUpstreamError(result.response.status, errText2)) {
+      const delayMs = replayBackoffMs(attempt);
+      onRetry?.({ attempt, status: result.response.status, detail: errText2, delayMs, maxAttempts });
+      await sleep(delayMs, externalSignal);
+      continue;
+    }
+    throw new UpstreamHttpError(result.response.status, errText2, attempt);
+  }
+}
+
 // src/upstream-proxy.ts
 var dispatcherCache = /* @__PURE__ */ new Map();
 var lastConnection = {};
@@ -47519,12 +47692,21 @@ function resolveProxyDecision(routes, globalProxy, upstreamUrl, fallback = {}) {
 function resolveProxy(routes, globalProxy, upstreamUrl, fallback = {}) {
   return resolveProxyDecision(routes, globalProxy, upstreamUrl, fallback).proxy;
 }
-function proxyDispatcher(proxyUrl) {
+function proxyDispatcher(proxyUrl, timeoutMs) {
   if (!proxyUrl) return void 0;
-  let agent = dispatcherCache.get(proxyUrl);
+  const t = timeoutMs ?? upstreamTimeoutMs();
+  const key = `${proxyUrl}\0${t}`;
+  let agent = dispatcherCache.get(key);
   if (!agent) {
-    agent = new import_undici.ProxyAgent({ uri: proxyUrl });
-    dispatcherCache.set(proxyUrl, agent);
+    const withTimeouts = (options) => ({ ...options, headersTimeout: t, bodyTimeout: t });
+    agent = new import_undici2.ProxyAgent({
+      uri: proxyUrl,
+      headersTimeout: t,
+      bodyTimeout: t,
+      factory: (origin, options) => new import_undici2.Pool(origin, withTimeouts(options)),
+      clientFactory: (origin, options) => new import_undici2.Pool(origin, withTimeouts(options))
+    });
+    dispatcherCache.set(key, agent);
   }
   return agent;
 }
@@ -47677,6 +47859,95 @@ function getUpstreamConnectionStatus() {
   return { ...lastConnection };
 }
 
+// src/compat-roles.ts
+function parseCompatRoles(v2) {
+  if (!v2 || typeof v2 !== "object" || Array.isArray(v2)) return void 0;
+  const obj = v2;
+  let out;
+  for (const [k2, val] of Object.entries(obj)) {
+    if (typeof val !== "string" || val.length === 0) continue;
+    out ??= {};
+    out[k2] = val;
+  }
+  return out;
+}
+function resolveCompatRoles(routes, upstreamUrl, globalRoles) {
+  const providerRoles = findRoute(routes, upstreamUrl)?.compat?.roles;
+  if (!globalRoles && !providerRoles) return {};
+  return { ...globalRoles, ...providerRoles };
+}
+function applyCompatRolesJson(parsed, protocol, roles) {
+  if (Object.keys(roles).length === 0) return 0;
+  const items = protocol === "openai" ? parsed.messages : parsed.input;
+  if (!Array.isArray(items)) return 0;
+  let rewritten = 0;
+  for (const item of items) {
+    if (!item || typeof item !== "object" || Array.isArray(item)) continue;
+    const msg = item;
+    if (protocol === "responses" && msg.type !== void 0 && msg.type !== "message") continue;
+    const role = msg.role;
+    if (typeof role !== "string") continue;
+    const mapped = roles[role];
+    if (mapped === void 0 || mapped === role) continue;
+    msg.role = mapped;
+    rewritten++;
+  }
+  return rewritten;
+}
+var ROLE_CAPTURE_STOPWORDS = /* @__PURE__ */ new Set([
+  "must",
+  "should",
+  "be",
+  "is",
+  "one",
+  "of",
+  "the",
+  "a",
+  "an",
+  "in",
+  "for",
+  "not",
+  "was",
+  "and",
+  "or",
+  "to",
+  "only",
+  "allowed",
+  "supported",
+  "valid",
+  "value",
+  "message",
+  "messages"
+]);
+var ROLE_REJECTION_PATTERNS = [
+  /(?:invalid|unknown|unsupported|unrecognized|unexpected)[a-z ]{0,24}?role\s*[:=]?\s*["'`]?([a-z][a-z0-9_-]{1,31})["'`]?/i,
+  /role\s*[:=]?\s*["'`]?([a-z][a-z0-9_-]{1,31})["'`]?\s+(?:is\s+)?not\s+(?:supported|allowed|recognized|valid|accepted)/i
+];
+function detectRoleRejection(status, bodyText) {
+  if (status !== 400) return null;
+  const head = bodyText.slice(0, 4096);
+  for (const re2 of ROLE_REJECTION_PATTERNS) {
+    const m2 = re2.exec(head);
+    if (!m2) continue;
+    const role = m2[1].toLowerCase();
+    if (ROLE_CAPTURE_STOPWORDS.has(role)) continue;
+    return { role };
+  }
+  return null;
+}
+function applyCompatRoles(body, protocol, roles) {
+  if (Object.keys(roles).length === 0) return { body, rewritten: 0 };
+  let parsed;
+  try {
+    parsed = JSON.parse(body);
+  } catch {
+    return { body, rewritten: 0 };
+  }
+  const rewritten = applyCompatRolesJson(parsed, protocol, roles);
+  if (rewritten === 0) return { body, rewritten: 0 };
+  return { body: JSON.stringify(parsed), rewritten };
+}
+
 // src/config.ts
 function safeReadJson(path18) {
   try {
@@ -47823,6 +48094,7 @@ function loadOptions(env = process.env) {
     promptCache: {
       routing: parsePromptCacheRouting(env.ACP_PROMPT_CACHE_ROUTING ?? fileConfig.promptCache?.routing)
     },
+    compat: { roles: parseCompatRoles(fileConfig.compat?.roles) ?? {} },
     sessionHeader: env.ACP_SESSION_HEADER ?? fileConfig.sessionHeader ?? "x-acp-session",
     log: env.ACP_LOG !== "0" && fileConfig.log !== false,
     debug: (env.ACP_DEBUG ?? (fileConfig.debug ? "1" : "0")) === "1",
@@ -47892,6 +48164,8 @@ function parseRouteEntry(v2) {
     if (typeof obj.proxy === "string") route.proxy = obj.proxy;
     if (obj.compressProtocol === "marker" || obj.compressProtocol === "tools") route.compressProtocol = obj.compressProtocol;
     if (obj.compress) route.compress = obj.compress;
+    const compatRoles = parseCompatRoles(obj.compat?.roles);
+    if (compatRoles) route.compat = { roles: compatRoles };
     return route;
   }
   if (v2 === null) return {};
@@ -48072,157 +48346,6 @@ import { readFile, writeFile, mkdir } from "fs/promises";
 import { existsSync as existsSync2, statSync as statSync2 } from "fs";
 import path3 from "path";
 
-// src/fetch-util.ts
-var MAX_REQUEST_BYTES = 100 * 1024 * 1024;
-var UPSTREAM_TIMEOUT_MS = 10 * 60 * 1e3;
-var liveUpstreamTimers = /* @__PURE__ */ new Set();
-async function fetchWithTimeout(url, opts, timeoutMs = UPSTREAM_TIMEOUT_MS, externalSignal) {
-  const controller = new AbortController();
-  const armTimer = () => {
-    const t = setTimeout(() => {
-      liveUpstreamTimers.delete(t);
-      controller.abort();
-    }, timeoutMs);
-    liveUpstreamTimers.add(t);
-    return t;
-  };
-  let timer3 = armTimer();
-  const rearm = () => {
-    clearTimeout(timer3);
-    liveUpstreamTimers.delete(timer3);
-    timer3 = armTimer();
-  };
-  let onExternalAbort = null;
-  if (externalSignal) {
-    if (externalSignal.aborted) controller.abort();
-    else {
-      onExternalAbort = () => controller.abort();
-      externalSignal.addEventListener("abort", onExternalAbort, { once: true });
-    }
-  }
-  const cleanup = () => {
-    clearTimeout(timer3);
-    liveUpstreamTimers.delete(timer3);
-    if (onExternalAbort && externalSignal) externalSignal.removeEventListener("abort", onExternalAbort);
-  };
-  try {
-    const finalOpts = { ...opts, signal: controller.signal };
-    const raw = await fetch(url, finalOpts);
-    if (raw.body) {
-      const wrapped = armIdleBody(raw.body, rearm);
-      return {
-        response: new Response(wrapped, {
-          status: raw.status,
-          statusText: raw.statusText,
-          headers: raw.headers
-        }),
-        clearTimer: cleanup
-      };
-    }
-    return { response: raw, clearTimer: cleanup };
-  } catch (e) {
-    cleanup();
-    throw e;
-  }
-}
-function armIdleBody(body, rearm) {
-  const reader = body.getReader();
-  return new ReadableStream({
-    async pull(controller) {
-      try {
-        const result = await reader.read();
-        if (result.done) {
-          controller.close();
-          return;
-        }
-        rearm();
-        controller.enqueue(result.value);
-      } catch (e) {
-        controller.error(e);
-      }
-    },
-    async cancel(reason) {
-      try {
-        await reader.cancel(reason);
-      } catch {
-      }
-    }
-  });
-}
-var UpstreamHttpError = class extends Error {
-  status;
-  body;
-  attempts;
-  constructor(status, body, attempts) {
-    super(`upstream error ${status}`);
-    this.name = "UpstreamHttpError";
-    this.status = status;
-    this.body = body;
-    this.attempts = attempts;
-  }
-};
-var TRANSIENT_BODY_MARKERS = [
-  "captcha",
-  "verify failed",
-  "risk control",
-  "\u98CE\u63A7",
-  "rate limit",
-  "too many requests",
-  "try again"
-];
-function isTransientUpstreamError(status, body) {
-  if (status === 429 || status >= 500) return true;
-  if (status < 400) return false;
-  const lower = body.toLowerCase();
-  return TRANSIENT_BODY_MARKERS.some((marker) => lower.includes(marker));
-}
-var REPLAY_MAX_ATTEMPTS = 3;
-function replayMaxAttempts() {
-  const raw = Number(process.env.BILI_REPLAY_RETRY_MAX);
-  return Number.isInteger(raw) && raw >= 1 ? raw : REPLAY_MAX_ATTEMPTS;
-}
-function replayBaseDelayMs() {
-  const raw = Number(process.env.BILI_REPLAY_RETRY_BASE_MS);
-  return Number.isFinite(raw) && raw >= 0 ? raw : 1500;
-}
-function maxShrinkPerCompress() {
-  const raw = Number(process.env.BILI_MAX_SHRINK_PER_COMPRESS);
-  return Number.isFinite(raw) && raw > 0 && raw <= 1 ? raw : void 0;
-}
-function replayBackoffMs(attempt) {
-  return replayBaseDelayMs() * 2 ** (attempt - 1);
-}
-function sleep(ms2, signal) {
-  if (ms2 <= 0 || signal?.aborted) return Promise.resolve();
-  return new Promise((resolve) => {
-    let timer3 = null;
-    const finish2 = () => {
-      if (timer3) clearTimeout(timer3);
-      if (signal) signal.removeEventListener("abort", finish2);
-      resolve();
-    };
-    timer3 = setTimeout(finish2, ms2);
-    if (signal) signal.addEventListener("abort", finish2, { once: true });
-  });
-}
-async function fetchWithRetry(url, opts, timeoutMs, externalSignal, onRetry) {
-  const maxAttempts = replayMaxAttempts();
-  for (let attempt = 1; ; attempt++) {
-    const result = await fetchWithTimeout(url, opts, timeoutMs, externalSignal);
-    if (result.response.ok) return result;
-    const errText2 = await result.response.text().catch(() => "upstream error");
-    result.clearTimer();
-    const lastAttempt = attempt >= maxAttempts;
-    if (!lastAttempt && isTransientUpstreamError(result.response.status, errText2)) {
-      const delayMs = replayBackoffMs(attempt);
-      onRetry?.({ attempt, status: result.response.status, detail: errText2, delayMs, maxAttempts });
-      await sleep(delayMs, externalSignal);
-      continue;
-    }
-    throw new UpstreamHttpError(result.response.status, errText2, attempt);
-  }
-}
-
 // src/registry-snapshot.json
 var registry_snapshot_default = { fetchedAt: "2026-08-24T10:39:50.602Z", count: 355, models: { "tencent/hy3-preview": { id: "tencent/hy3-preview", name: "Hy3 preview", description: "Tencent Hy reasoning model for coding, instruction following, and agent tasks", family: "Hy", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-04-20", last_updated: "2026-04-20", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 64e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/tencent/Hy3-preview" }], benchmarks: [{ name: "SWE-Bench Verified", score: 74.4, metric: "resolved", source: "https://huggingface.co/tencent/Hy3-preview" }] }, "tencent/hy3": { id: "tencent/hy3", name: "Hy3", description: "Tencent Hy reasoning model for coding, instruction following, and agent tasks", family: "Hy", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-07-06", last_updated: "2026-07-06", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 64e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/tencent/Hy3" }], benchmarks: [{ name: "SWE-Bench Verified", score: 78, metric: "resolved", source: "https://huggingface.co/tencent/Hy3" }] }, "nvidia/llama-3.3-nemotron-super-49b-v1": { id: "nvidia/llama-3.3-nemotron-super-49b-v1", name: "Llama 3.3 Nemotron Super 49B v1", description: "Nemotron model for efficient reasoning, coding, and specialized AI agents", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-04-07", last_updated: "2025-04-07", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 } }, "nvidia/nemotron-3-nano-30b-a3b": { id: "nvidia/nemotron-3-nano-30b-a3b", name: "Nemotron 3 Nano 30B A3B", description: "Small Nemotron 3 MoE for efficient coding, math, and long-context agents", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-12-15", last_updated: "2025-12-15", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 } }, "nvidia/nemotron-3.5-content-safety": { id: "nvidia/nemotron-3.5-content-safety", name: "Nemotron 3.5 Content Safety", description: "Safety model for policy screening, moderation, and risk-aware routing workflows", family: "nemotron", attachment: true, reasoning: true, tool_call: false, temperature: true, release_date: "2026-06-04", last_updated: "2026-06-04", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8192 } }, "nvidia/nemotron-cascade-2-30b-a3b": { id: "nvidia/nemotron-cascade-2-30b-a3b", name: "Nemotron Cascade 2 30B A3B", description: "Nemotron model for efficient reasoning, coding, and specialized AI agents", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-03-24", last_updated: "2026-04-09", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 32768 } }, "nvidia/nemotron-3-content-safety": { id: "nvidia/nemotron-3-content-safety", name: "Nemotron 3 Content Safety", description: "Safety model for policy screening, moderation, and risk-aware routing workflows", family: "nemotron", attachment: false, reasoning: false, tool_call: false, temperature: false, release_date: "2026-04-16", last_updated: "2026-04-16", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 } }, "nvidia/llama-3.1-nemotron-70b-instruct": { id: "nvidia/llama-3.1-nemotron-70b-instruct", name: "Llama 3.1 Nemotron 70B Instruct", description: "Nemotron model for efficient reasoning, coding, and specialized AI agents", family: "nemotron", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2025-04-15", last_updated: "2025-04-15", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8192 } }, "nvidia/nemotron-nano-12b-v2-vl": { id: "nvidia/nemotron-nano-12b-v2-vl", name: "Nemotron Nano 12B v2 VL", description: "Nemotron multimodal model for visual reasoning and agentic AI workflows", family: "nemotron", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2025-10-28", last_updated: "2025-10-28", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 } }, "nvidia/nemotron-voicechat": { id: "nvidia/nemotron-voicechat", name: "Nemotron VoiceChat", description: "Nemotron multimodal model for visual reasoning and agentic AI workflows", family: "nemotron", attachment: true, reasoning: false, tool_call: true, temperature: true, release_date: "2026-03-16", last_updated: "2026-03-16", modalities: { input: ["text", "audio"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8192 } }, "nvidia/llama-nemotron-rerank-vl-1b-v2": { id: "nvidia/llama-nemotron-rerank-vl-1b-v2", name: "Llama Nemotron Rerank VL 1B v2", description: "Reranking model for improving retrieval quality in search and recommendation systems", family: "nemotron", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2026-03-31", last_updated: "2026-03-31", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 } }, "nvidia/llama-3.3-nemotron-super-49b-v1.5": { id: "nvidia/llama-3.3-nemotron-super-49b-v1.5", name: "Llama 3.3 Nemotron Super 49B v1.5", description: "Nemotron model for efficient reasoning, coding, and specialized AI agents", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-07-25", last_updated: "2025-07-25", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 } }, "nvidia/nemotron-3-super-120b-a12b": { id: "nvidia/nemotron-3-super-120b-a12b", name: "Nemotron 3 Super 120B A12B", description: "Nemotron middle tier for collaborative agents and high-volume reasoning workloads", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-03-11", last_updated: "2026-03-11", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 } }, "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning": { id: "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", name: "Nemotron 3 Nano Omni 30B A3B Reasoning", description: "Open Nemotron omni model combining reasoning with text, vision, and audio", family: "nemotron", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-04-28", last_updated: "2026-04-28", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 65536 } }, "nvidia/nemotron-3.5-lightning": { id: "nvidia/nemotron-3.5-lightning", name: "Nemotron 3.5 Lightning 30B A3B", description: "Fast NVIDIA Nemotron MoE for reliable agentic tasks across enterprise workloads", family: "nemotron", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-11", last_updated: "2026-08-11", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 } }, "nvidia/nemotron-content-safety-reasoning-4b": { id: "nvidia/nemotron-content-safety-reasoning-4b", name: "Nemotron Content Safety Reasoning 4B", description: "Safety model for policy screening, moderation, and risk-aware routing workflows", family: "nemotron", attachment: false, reasoning: true, tool_call: false, temperature: false, release_date: "2026-01-22", last_updated: "2026-01-22", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 } }, "nvidia/nemotron-3-ultra-550b-a55b": { id: "nvidia/nemotron-3-ultra-550b-a55b", name: "Nemotron 3 Ultra 550B A55B", description: "Largest Nemotron 3 model for maximum open-weight reasoning and agent accuracy", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-04", last_updated: "2026-06-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Verified", score: 70.7, metric: "resolved", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "SWE-Bench Multilingual", score: 67.7, metric: "resolve rate", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "Terminal-Bench", score: 56.4, metric: "success rate", version: "2.1", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "GPQA", score: 87, metric: "accuracy", variant: "no tools", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "Humanity's Last Exam", score: 26.7, metric: "accuracy", variant: "no tools", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "Humanity's Last Exam", score: 37.4, metric: "accuracy", variant: "with tools", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "LiveCodeBench", score: 89, metric: "pass@1", version: "v6", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "MMLU-Pro", score: 86.8, metric: "accuracy", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "BrowseComp", score: 44.4, metric: "accuracy", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "IFBench", score: 81.7, metric: "accuracy", variant: "prompt loose", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }, { name: "GDPval", score: 46.7, metric: "wins or ties", source: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16", date: "2026-06-04" }] }, "nvidia/mistral-nemotron": { id: "nvidia/mistral-nemotron", name: "Mistral Nemotron", description: "Mistral model for multilingual chat, reasoning, and tool-assisted workflows", family: "nemotron", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2025-06-11", last_updated: "2025-06-12", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8192 } }, "nvidia/llama-3.1-nemotron-safety-guard-8b-v3": { id: "nvidia/llama-3.1-nemotron-safety-guard-8b-v3", name: "Llama 3.1 Nemotron Safety Guard 8B v3", description: "Safety model for policy screening, moderation, and risk-aware routing workflows", family: "nemotron", attachment: false, reasoning: false, tool_call: false, temperature: false, release_date: "2025-10-28", last_updated: "2025-10-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 } }, "nvidia/nemotron-nano-9b-v2": { id: "nvidia/nemotron-nano-9b-v2", name: "Nemotron Nano 9B v2", description: "Compact Nemotron model for efficient reasoning and deployable AI agents", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-08-18", last_updated: "2025-08-18", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 } }, "nvidia/llama-3.1-nemotron-ultra-253b": { id: "nvidia/llama-3.1-nemotron-ultra-253b", name: "Llama 3.1 Nemotron Ultra 253B", description: "Flagship Nemotron model for high-throughput reasoning and complex agents", family: "nemotron", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-04-07", last_updated: "2025-04-07", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8192 } }, "nvidia/nemotron-mini-4b-instruct": { id: "nvidia/nemotron-mini-4b-instruct", name: "Nemotron Mini 4B Instruct", description: "Compact Nemotron model for efficient reasoning and deployable AI agents", family: "nemotron", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2024-08-21", last_updated: "2024-08-26", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8192 } }, "nvidia/llama-nemotron-embed-vl-1b-v2": { id: "nvidia/llama-nemotron-embed-vl-1b-v2", name: "Llama Nemotron Embed VL 1B v2", description: "Embedding model for semantic search, retrieval, clustering, and ranking pipelines", family: "nemotron", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2026-02-10", last_updated: "2026-02-10", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 32768, output: 2048 } }, "aisingapore/gemma-sea-lion-v4-27b-it": { id: "aisingapore/gemma-sea-lion-v4-27b-it", name: "Gemma-SEA-LION-v4-27B-IT", description: "Gemma 3 27B tuned by AI Singapore for Southeast Asian languages and instruction following", family: "gemma", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2025-09-23", last_updated: "2025-09-23", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/aisingapore/Gemma-SEA-LION-v4-27B-IT" }] }, "microsoft/phi-4-mini": { id: "microsoft/phi-4-mini", name: "Phi-4-mini", description: "Compact Microsoft instruction model tuned for efficient coding assistance, reasoning, and low-latency agent tasks", family: "phi", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2023-10", release_date: "2024-12-11", last_updated: "2024-12-11", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 }, links: [{ label: "Weights", url: "https://huggingface.co/microsoft/Phi-4-mini-instruct", type: "weights" }], benchmarks: [{ name: "MMLU", score: 67.3, metric: "accuracy", source: "https://huggingface.co/microsoft/Phi-4-mini-instruct/resolve/main/README.md" }] }, "microsoft/mai-code-1-flash": { id: "microsoft/mai-code-1-flash", name: "MAI-Code-1-Flash", description: "Microsoft coding model built for fast, efficient assistance in everyday developer workflows", family: "mai", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-12", release_date: "2026-06-02", last_updated: "2026-06-08", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 128e3 }, links: [{ label: "Model card", url: "https://microsoft.ai/pdf/MAI-Code-1-Flash-Model-Card.PDF", type: "model_card" }, { label: "Announcement", url: "https://microsoft.ai/news/introducingmai-code-1-flash/", type: "announcement" }], benchmarks: [{ name: "SWE-Bench Pro", score: 51.2, metric: "resolve rate", harness: "GitHub Copilot", source: "https://microsoft.ai/news/introducingmai-code-1-flash/", date: "2026-06-02" }, { name: "SWE-Bench Verified", score: 71.6, metric: "resolved", source: "https://llm-stats.com/benchmarks/swe-bench-verified" }, { name: "Terminal-Bench", score: 54.8, metric: "success rate", version: "2.0", source: "https://llm-stats.com/benchmarks/terminal-bench-2" }, { name: "GPQA Diamond", score: 84.6, metric: "accuracy", source: "https://llm-stats.com/benchmarks/gpqa" }] }, "microsoft/mai-code-1.1-flash": { id: "microsoft/mai-code-1.1-flash", name: "MAI-Code-1.1-Flash", description: "Microsoft coding model with native vision support, optimized for fast and efficient software development", family: "mai", attachment: true, reasoning: true, tool_call: true, structured_output: true, release_date: "2026-08-11", last_updated: "2026-08-11", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 128e3 }, links: [{ label: "Announcement", url: "https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/", type: "announcement" }] }, "deepseek/deepseek-r1-distill-qwen-32b": { id: "deepseek/deepseek-r1-distill-qwen-32b", name: "DeepSeek-R1-Distill-Qwen-32B", description: "R1 reasoning distilled into Qwen 2.5 32B for efficient open-weight step-by-step problem solving", family: "deepseek-thinking", attachment: false, reasoning: true, tool_call: false, temperature: true, release_date: "2025-01-20", last_updated: "2025-01-20", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }] }, "deepseek/deepseek-r1": { id: "deepseek/deepseek-r1", name: "DeepSeek-R1", description: "Classic open reasoning model for transparent math, coding, and deliberate problem solving", family: "deepseek-thinking", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-07", release_date: "2025-01-20", last_updated: "2025-05-29", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-R1" }], benchmarks: [{ name: "Aider Polyglot", score: 56.9, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-01-20" }, { name: "Artificial Analysis Coding Index", score: 15.9, metric: "index", source: "https://openrouter.ai/deepseek/deepseek-r1/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 35.7, metric: "percent correct", source: "https://openrouter.ai/deepseek/deepseek-r1/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 6.1, metric: "success rate", source: "https://openrouter.ai/deepseek/deepseek-r1/benchmarks", date: "2026-03-11" }] }, "deepseek/deepseek-v3-0324": { id: "deepseek/deepseek-v3-0324", name: "DeepSeek V3 0324", description: "March 2025 checkpoint of DeepSeek-V3 with improved reasoning and coding", family: "deepseek", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2025-03-24", last_updated: "2025-03-24", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 163840, output: 163840 }, weights: [{ label: "Model weights", url: "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324", format: "safetensors" }] }, "deepseek/deepseek-v4-flash": { id: "deepseek/deepseek-v4-flash", name: "DeepSeek V4 Flash", description: "Fast DeepSeek V4 lane for economical reasoning, coding, and long-context work", family: "deepseek-flash", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-05", release_date: "2026-04-24", last_updated: "2026-04-24", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash" }], benchmarks: [{ name: "SWE-Bench Verified", score: 79, metric: "resolved", source: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash" }] }, "deepseek/deepseek-v4-pro-0813": { id: "deepseek/deepseek-v4-pro-0813", name: "DeepSeek V4 Pro 0813", description: "DeepSeek V4 Pro snapshot with million-token context and support for thinking and non-thinking modes", family: "deepseek-thinking", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-12", last_updated: "2026-08-22", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 }, license: "MIT", weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813" }] }, "deepseek/deepseek-v3.1": { id: "deepseek/deepseek-v3.1", name: "DeepSeek-V3.1", description: "Hybrid-reasoning DeepSeek model with thinking and non-thinking modes", family: "deepseek", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-08-21", last_updated: "2025-08-21", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, license: "MIT License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V3.1" }] }, "deepseek/deepseek-v4-pro": { id: "deepseek/deepseek-v4-pro", name: "DeepSeek V4 Pro", description: "Open MoE flagship with million-token context for coding and long agent runs", family: "deepseek-thinking", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-05", release_date: "2026-04-24", last_updated: "2026-04-24", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro" }], benchmarks: [{ name: "SWE-Bench Verified", score: 80.6, metric: "resolved", source: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro" }, { name: "Artificial Analysis Coding Agent Index", score: 50.1, metric: "average pass@1", harness: "Claude Code", variant: "high", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 67.8, metric: "pass@1", harness: "Claude Code", variant: "high", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 18, metric: "pass@1", harness: "Claude Code", variant: "high", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 64.7, metric: "pass@1", harness: "Claude Code", variant: "high", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }] }, "deepseek/deepseek-v3": { id: "deepseek/deepseek-v3", name: "DeepSeek-V3", description: "Open DeepSeek MoE chat model for coding, math, and general reasoning", family: "deepseek", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2024-12-26", last_updated: "2024-12-26", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, license: "DeepSeek Model License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V3" }] }, "deepseek/deepseek-v3.2": { id: "deepseek/deepseek-v3.2", name: "DeepSeek V3.2", description: "Hybrid-reasoning DeepSeek model with thinking and non-thinking modes, sparse attention, and tool-use", family: "deepseek", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-07", release_date: "2025-12-01", last_updated: "2025-12-01", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 64e3 }, license: "MIT License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V3.2" }] }, "deepseek/deepseek-chat": { id: "deepseek/deepseek-chat", name: "DeepSeek Chat", description: "DeepSeek chat model for instruction following, coding, and analysis", family: "deepseek", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-09", release_date: "2025-12-01", last_updated: "2026-02-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V3.2" }], benchmarks: [{ name: "Aider Polyglot", score: 70.2, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-10-03" }] }, "deepseek/deepseek-reasoner": { id: "deepseek/deepseek-reasoner", name: "DeepSeek Reasoner", description: "DeepSeek reasoning model for multi-step analysis, math, coding, and tools", family: "deepseek-thinking", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-09", release_date: "2025-12-01", last_updated: "2026-02-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V3.2" }], benchmarks: [{ name: "Aider Polyglot", score: 74.2, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-10-03" }] }, "deepseek/deepseek-ocr-2": { id: "deepseek/deepseek-ocr-2", name: "DeepSeek OCR 2", description: "High-accuracy OCR model for extracting text from documents, screenshots, receipts, and natural scenes", attachment: true, reasoning: false, tool_call: false, release_date: "2026-01-27", last_updated: "2026-01-27", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 8192, output: 8192 } }, "deepseek/deepseek-v4-flash-0731": { id: "deepseek/deepseek-v4-flash-0731", name: "DeepSeek V4 Flash 0731", description: "Official DeepSeek V4 Flash release with enhanced agentic capabilities and integrated DSpark speculative decoding", family: "deepseek-flash", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-05", release_date: "2026-07-31", last_updated: "2026-07-31", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 }, license: "MIT", weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731" }], benchmarks: [{ name: "Terminal-Bench", score: 82.7, metric: "pass@1", variant: "max", version: "2.1", source: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731" }, { name: "NL2Repo", score: 54.2, metric: "resolve rate", harness: "DeepSeek Harness minimal mode", variant: "max effort", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "CyberGym", score: 76.7, metric: "score", harness: "DeepSeek Harness minimal mode", variant: "max effort", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "DeepSWE", score: 54.4, metric: "resolve rate", harness: "DeepSeek Harness minimal mode", variant: "max effort", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "Toolathlon-Verified", score: 70.3, metric: "score", harness: "DeepSeek Harness minimal mode", variant: "max effort", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "Agents' Last Exam", score: 25.2, metric: "score", harness: "DeepSeek Harness minimal mode", variant: "max effort", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "AutomationBench", score: 25.1, metric: "success rate", variant: "max effort", dataset: "public", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "DSBench-FullStack", score: 68.7, metric: "score", variant: "max effort", dataset: "internal", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }, { name: "DSBench-Hard", score: 59.6, metric: "score", variant: "max effort", dataset: "internal", source: "https://api-docs.deepseek.com/updates/", date: "2026-07-31" }] }, "deepseek/deepseek-v4-pro-0423": { id: "deepseek/deepseek-v4-pro-0423", name: "DeepSeek V4 Pro 0423", description: "DeepSeek V4 Pro initial snapshot with million-token context and support for thinking and non-thinking modes", family: "deepseek-thinking", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-05", release_date: "2026-04-23", last_updated: "2026-04-23", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 384e3 } }, "deepseek/deepseek-v4-flash-vision-exp": { id: "deepseek/deepseek-v4-flash-vision-exp", name: "DeepSeek V4 Flash Vision Exp", description: "Experimental multimodal DeepSeek V4 Flash model for image understanding, coding, and agentic work", family: "deepseek-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-21", last_updated: "2026-08-21", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 384e3 } }, "arcee-ai/trinity-large-preview": { id: "arcee-ai/trinity-large-preview", name: "Trinity Large Preview", description: "Lightly post-trained 398B MoE chat model for creative work, long-context prompts, and tool-using agents", family: "trinity", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2026-01-27", last_updated: "2026-05-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 524288, output: 262144 }, license: "OpenMDW-1.1", links: [{ label: "Model card", url: "https://huggingface.co/arcee-ai/Trinity-Large-Preview", type: "model_card" }, { label: "Announcement", url: "https://www.arcee.ai/blog/trinity-large", type: "announcement" }, { label: "License", url: "https://huggingface.co/arcee-ai/Trinity-Large-Preview/blob/main/LICENSE", type: "license" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/arcee-ai/Trinity-Large-Preview", format: "safetensors" }] }, "arcee-ai/trinity-mini": { id: "arcee-ai/trinity-mini", name: "Trinity Mini", description: "Reasoning-tuned 26B MoE model with 3B active parameters for agents, tools, and multi-step workloads", family: "trinity", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-12-01", last_updated: "2026-05-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 }, license: "OpenMDW-1.1", links: [{ label: "Model card", url: "https://huggingface.co/arcee-ai/Trinity-Mini", type: "model_card" }, { label: "Announcement", url: "https://www.arcee.ai/blog/the-trinity-manifesto", type: "announcement" }, { label: "License", url: "https://huggingface.co/arcee-ai/Trinity-Mini/blob/main/LICENSE", type: "license" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/arcee-ai/Trinity-Mini", format: "safetensors" }] }, "arcee-ai/trinity-large-thinking": { id: "arcee-ai/trinity-large-thinking", name: "Trinity Large Thinking", description: "Reasoning-optimized 398B MoE agent model with extended thinking for long-horizon and multi-turn tool use", family: "trinity", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-04-01", last_updated: "2026-05-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 524288, output: 262144 }, license: "OpenMDW-1.1", links: [{ label: "Model card", url: "https://huggingface.co/arcee-ai/Trinity-Large-Thinking", type: "model_card" }, { label: "Announcement", url: "https://www.arcee.ai/blog/trinity-large-thinking", type: "announcement" }, { label: "License", url: "https://huggingface.co/arcee-ai/Trinity-Large-Thinking/blob/main/LICENSE", type: "license" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/arcee-ai/Trinity-Large-Thinking", format: "safetensors" }] }, "arcee-ai/trinity-nano-preview": { id: "arcee-ai/trinity-nano-preview", name: "Trinity Nano Preview", description: "Experimental chat-tuned 6B MoE model with 1B active parameters for low-resource chat and instruction following", family: "trinity", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2025-12-01", last_updated: "2026-05-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 }, license: "OpenMDW-1.1", links: [{ label: "Model card", url: "https://huggingface.co/arcee-ai/Trinity-Nano-Preview", type: "model_card" }, { label: "Announcement", url: "https://www.arcee.ai/blog/the-trinity-manifesto", type: "announcement" }, { label: "License", url: "https://huggingface.co/arcee-ai/Trinity-Nano-Preview/blob/main/LICENSE", type: "license" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/arcee-ai/Trinity-Nano-Preview", format: "safetensors" }] }, "google/gemini-3.1-flash-tts-preview": { id: "google/gemini-3.1-flash-tts-preview", name: "Gemini 3.1 Flash TTS Preview", description: "Low-latency speech generation with steerable prompts and expressive audio tags", family: "gemini-flash", attachment: false, reasoning: false, tool_call: false, temperature: true, knowledge: "2025-01", release_date: "2026-04-15", last_updated: "2026-04-15", modalities: { input: ["text"], output: ["audio"] }, open_weights: false, limit: { context: 8192, output: 16384 } }, "google/gemma-4-26b-a4b-it": { id: "google/gemma-4-26b-a4b-it", name: "Gemma 4 26B A4B IT", description: "Open Gemma instruction model for efficient chat and self-hosted deployments", family: "gemma", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-02", last_updated: "2026-04-02", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/google/gemma-4-26B-A4B-it" }] }, "google/gemini-3-pro-image": { id: "google/gemini-3-pro-image", name: "Nano Banana Pro", description: "Nano Banana Pro for higher-fidelity image generation and design-heavy edits", family: "gemini-pro", attachment: true, reasoning: true, tool_call: false, temperature: true, knowledge: "2025-01", release_date: "2026-05-28", last_updated: "2026-05-28", modalities: { input: ["text", "image"], output: ["text", "image"] }, open_weights: false, limit: { context: 65536, output: 32768 } }, "google/gemini-3-pro-image-preview": { id: "google/gemini-3-pro-image-preview", name: "Nano Banana Pro", description: "Nano Banana Pro for higher-fidelity image generation and design-heavy edits", family: "gemini-pro", attachment: true, reasoning: true, tool_call: false, temperature: true, knowledge: "2025-01", release_date: "2025-11-20", last_updated: "2025-11-20", modalities: { input: ["text", "image"], output: ["text", "image"] }, open_weights: false, limit: { context: 65536, output: 32768 } }, "google/gemini-flash-lite-latest": { id: "google/gemini-flash-lite-latest", name: "Gemini Flash-Lite Latest", description: "Fast Gemini model balancing multimodal reasoning, tool use, and cost", family: "gemini-flash-lite", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-03", release_date: "2026-07-21", last_updated: "2026-07-21", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/gemini-2.0-flash": { id: "google/gemini-2.0-flash", name: "Gemini 2.0 Flash", description: "Earlier Gemini Flash workhorse for responsive multimodal apps and tool use", family: "gemini-flash", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-06", release_date: "2024-12-11", last_updated: "2024-12-11", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 8192 } }, "google/gemini-2.5-flash-tts": { id: "google/gemini-2.5-flash-tts", name: "Gemini 2.5 Flash TTS", description: "Speech generation model for controllable voice, narration, and audio delivery", family: "gemini-flash", attachment: false, reasoning: false, tool_call: false, temperature: true, knowledge: "2025-01", release_date: "2025-09-30", last_updated: "2025-12-10", modalities: { input: ["text"], output: ["audio"] }, open_weights: false, limit: { context: 32768, output: 16384 } }, "google/gemini-3.5-flash-lite": { id: "google/gemini-3.5-flash-lite", name: "Gemini 3.5 Flash Lite", description: "Fast Gemini model balancing multimodal reasoning, tool use, and cost", family: "gemini-flash-lite", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-03", release_date: "2026-07-21", last_updated: "2026-07-21", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "SWE-Bench Pro", score: 54.2, metric: "resolve rate", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "Terminal-Bench", score: 54, metric: "accuracy", harness: "Terminus 2", version: "2.1", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "MLE-Bench", score: 39.2, metric: "average position score", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "GDPval-AA", score: 1140, metric: "Elo", version: "v2", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "OSWorld-Verified", score: 74, metric: "success rate", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "CharXiv Reasoning", score: 74.5, metric: "accuracy", variant: "no tools", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "CharXiv Reasoning", score: 76.5, metric: "accuracy", variant: "with tools", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "GDM-MRCR", score: 72.2, metric: "accuracy", variant: "128k average, 8-needle", version: "v2", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }, { name: "GDM-MRCR", score: 21.3, metric: "accuracy", variant: "1M pointwise, 8-needle", version: "v2", source: "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/", date: "2026-07-21" }] }, "google/gemini-3.1-flash-lite-image": { id: "google/gemini-3.1-flash-lite-image", name: "Nano Banana 2 Lite", description: "Fastest, most cost-efficient Gemini image model for high-volume 1K generation and editing", family: "gemini-flash-lite", attachment: true, reasoning: true, tool_call: true, structured_output: false, temperature: true, knowledge: "2025-01", release_date: "2026-06-30", last_updated: "2026-06-30", modalities: { input: ["text", "image"], output: ["text", "image"] }, open_weights: false, limit: { context: 65536, output: 4096 } }, "google/gemini-3.1-pro-preview": { id: "google/gemini-3.1-pro-preview", name: "Gemini 3.1 Pro Preview", description: "Reasoning-first Gemini preview for agentic coding and complex problem solving", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-02-19", last_updated: "2026-02-19", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "SWE-Bench Pro", score: 54.2, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "Terminal-Bench", score: 70.3, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "SWE-Bench Pro", score: 46.1, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }, { name: "SWE-Atlas Codebase QnA", score: 13.5, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 33.81, metric: "score", harness: "Gemini CLI", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 29.84, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "Artificial Analysis Coding Agent Index", score: 43, metric: "average pass@1", harness: "Gemini CLI", variant: "high", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 45.6, metric: "pass@1", harness: "Gemini CLI", variant: "high", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 15.1, metric: "pass@1", harness: "Gemini CLI", variant: "high", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 68.3, metric: "pass@1", harness: "Gemini CLI", variant: "high", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "GPQA Diamond", score: 94.3, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 44.4, metric: "accuracy", dataset: "full set, text + MM", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "ARC-AGI-2", score: 77.1, metric: "accuracy", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "MMMU Pro", score: 80.5, metric: "accuracy", variant: "no tools", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "MCP Atlas", score: 78.2, metric: "success rate", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "OSWorld-Verified", score: 76.2, metric: "success rate", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "CharXiv Reasoning", score: 83.3, metric: "accuracy", variant: "no tools", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "GDPval-AA", score: 1314, metric: "Elo", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }] }, "google/gemini-3.1-pro-preview-customtools": { id: "google/gemini-3.1-pro-preview-customtools", name: "Gemini 3.1 Pro Preview Custom Tools", description: "Advanced Gemini model for complex reasoning, coding, and multimodal analysis", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-02-19", last_updated: "2026-02-19", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/deep-research-preview-04-2026": { id: "google/deep-research-preview-04-2026", name: "Gemini Deep Research Preview", description: "Agentic model for autonomous multi-step research, synthesis, and cited reports", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2025-01", release_date: "2026-04-21", last_updated: "2026-04-21", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text", "image"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/gemini-2.5-flash-lite": { id: "google/gemini-2.5-flash-lite", name: "Gemini 2.5 Flash-Lite", description: "Lean Gemini 2.5 lane for cheap multimodal traffic and quick agents", family: "gemini-flash-lite", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2025-06-17", last_updated: "2025-06-17", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 9.5, metric: "index", source: "https://openrouter.ai/google/gemini-2.5-flash-lite/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 19.3, metric: "percent correct", source: "https://openrouter.ai/google/gemini-2.5-flash-lite/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 4.5, metric: "success rate", source: "https://openrouter.ai/google/gemini-2.5-flash-lite/benchmarks", date: "2026-03-11" }] }, "google/gemini-robotics-er-1.6-preview": { id: "google/gemini-robotics-er-1.6-preview", name: "Gemini Robotics-ER 1.6 Preview", description: "Vision-language model for embodied reasoning: spatial understanding, task planning, and physical-world agentic robotics", family: "gemini", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-04-14", last_updated: "2026-04-14", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 65536 } }, "google/gemma-4-31b-it": { id: "google/gemma-4-31b-it", name: "Gemma 4 31B IT", description: "Largest Gemma 4 instruction model for open, self-hosted chat and reasoning", family: "gemma", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-02", last_updated: "2026-04-02", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/google/gemma-4-31B-it" }] }, "google/gemini-2.5-computer-use-preview-10-2025": { id: "google/gemini-2.5-computer-use-preview-10-2025", name: "Gemini 2.5 Computer Use Preview", description: "Specialized Gemini 2.5 model for browser-control agents that automate UI tasks", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-01", release_date: "2025-10-07", last_updated: "2025-10-07", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 64e3 } }, "google/lyria-3-clip-preview": { id: "google/lyria-3-clip-preview", name: "Lyria 3 Clip Preview", description: "Music generation model for short 30-second clips, loops, and previews from text or image prompts", family: "lyria", attachment: true, reasoning: false, tool_call: false, structured_output: false, temperature: true, release_date: "2026-03-25", last_updated: "2026-03-25", modalities: { input: ["text", "image"], output: ["text", "audio"] }, open_weights: false, limit: { context: 131072, output: 65536 } }, "google/veo-3.1-fast-generate-preview": { id: "google/veo-3.1-fast-generate-preview", name: "Veo 3.1 Fast Preview", description: "Video model for prompt-guided generation, editing, and motion workflows", family: "veo", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2025-10-15", last_updated: "2026-01-01", modalities: { input: ["text", "image", "video"], output: ["video"] }, open_weights: false, limit: { context: 1024, output: 0 } }, "google/gemini-embedding-001": { id: "google/gemini-embedding-001", name: "Gemini Embedding 001", description: "Embedding model for semantic search, retrieval, clustering, and ranking pipelines", family: "gemini", attachment: false, reasoning: false, tool_call: false, temperature: false, knowledge: "2025-05", release_date: "2025-05-20", last_updated: "2025-05-20", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 2048, output: 1 } }, "google/gemini-flash-latest": { id: "google/gemini-flash-latest", name: "Gemini Flash Latest", description: "High-efficiency Gemini model for agentic workflows, coding, and multimodal reasoning", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-03", release_date: "2026-08-13", last_updated: "2026-08-13", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/gemini-3.5-flash": { id: "google/gemini-3.5-flash", name: "Gemini 3.5 Flash", description: "Fast Gemini model balancing multimodal reasoning, tool use, and cost", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-05-19", last_updated: "2026-05-19", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "Terminal-Bench", score: 76.2, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "SWE-Bench Pro", score: 55.1, metric: "resolve rate", variant: "single attempt", dataset: "public", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "MCP Atlas", score: 83.6, metric: "success rate", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "Toolathlon", score: 56.5, metric: "success rate", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "OSWorld-Verified", score: 78.4, metric: "success rate", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "MMMU Pro", score: 83.6, metric: "accuracy", variant: "no tools", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "CharXiv Reasoning", score: 84.2, metric: "accuracy", variant: "no tools", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "Humanity's Last Exam", score: 40.2, metric: "accuracy", dataset: "full set, text + MM", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "ARC-AGI-2", score: 72.1, metric: "accuracy", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }, { name: "GDPval-AA", score: 1656, metric: "Elo", source: "https://deepmind.google/models/gemini/flash/", date: "2026-05-19" }] }, "google/gemini-3.5-live-translate-preview": { id: "google/gemini-3.5-live-translate-preview", name: "Gemini 3.5 Live Translate Preview", description: "Low-latency audio-to-audio model for real-time speech translation across 70+ languages", family: "gemini-pro", attachment: false, reasoning: false, tool_call: false, temperature: false, knowledge: "2025-01", release_date: "2026-06-09", last_updated: "2026-06-09", modalities: { input: ["audio"], output: ["audio", "text"] }, open_weights: false, limit: { context: 131072, output: 65536 } }, "google/veo-3.1-lite-generate-preview": { id: "google/veo-3.1-lite-generate-preview", name: "Veo 3.1 Lite Preview", description: "Video model for prompt-guided generation, editing, and motion workflows", family: "veo", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2026-03-31", last_updated: "2026-03-31", modalities: { input: ["text", "image"], output: ["video"] }, open_weights: false, limit: { context: 1024, output: 0 } }, "google/gemini-omni-flash-preview": { id: "google/gemini-omni-flash-preview", name: "Gemini Omni Flash Preview", description: "Video generation and editing model for fast, conversational text- and image-to-video workflows", family: "gemini", attachment: true, reasoning: true, tool_call: false, release_date: "2026-06-30", last_updated: "2026-06-30", modalities: { input: ["text", "image", "video"], output: ["video"] }, open_weights: false, limit: { context: 1048576, output: 57920 }, benchmarks: [{ name: "LMArena Text-to-Video Arena", score: 1527, metric: "Elo", source: "https://venturebeat.com/technology/googles-gemini-omni-flash-hits-the-api-turning-enterprise-video-production-into-a-conversation", date: "2026-06-30" }] }, "google/gemini-3.1-flash-lite": { id: "google/gemini-3.1-flash-lite", name: "Gemini 3.1 Flash Lite", description: "Low-latency Gemini model for high-volume multimodal and agent workloads", family: "gemini-flash-lite", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-05-07", last_updated: "2026-05-07", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/gemini-2.5-flash": { id: "google/gemini-2.5-flash", name: "Gemini 2.5 Flash", description: "Fast Gemini workhorse for multimodal apps where latency and price matter", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2025-06-17", last_updated: "2025-06-17", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "Aider Polyglot", score: 55.1, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-25" }, { name: "Artificial Analysis Coding Index", score: 22.2, metric: "index", source: "https://openrouter.ai/google/gemini-2.5-flash/benchmarks", date: "2026-06-02" }, { name: "SciCode", score: 39.4, metric: "percent correct", source: "https://openrouter.ai/google/gemini-2.5-flash/benchmarks", date: "2026-06-02" }, { name: "Terminal-Bench Hard", score: 13.6, metric: "success rate", source: "https://openrouter.ai/google/gemini-2.5-flash/benchmarks", date: "2026-06-02" }] }, "google/gemini-3-pro-preview": { id: "google/gemini-3-pro-preview", name: "Gemini 3 Pro Preview", description: "Preview Gemini flagship for complex reasoning, coding, and rich multimodal prompts", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2025-11-18", last_updated: "2025-11-18", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "SWE-Bench Pro", score: 43.3, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "google/veo-3.1-generate-preview": { id: "google/veo-3.1-generate-preview", name: "Veo 3.1 Preview", description: "Video model for prompt-guided generation, editing, and motion workflows", family: "veo", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2025-10-15", last_updated: "2026-01", modalities: { input: ["text", "image"], output: ["video"] }, open_weights: false, limit: { context: 1024, output: 1 } }, "google/gemini-2.5-pro-tts": { id: "google/gemini-2.5-pro-tts", name: "Gemini 2.5 Pro TTS", description: "Speech generation model for controllable voice, narration, and audio delivery", family: "gemini-pro", attachment: false, reasoning: false, tool_call: false, temperature: false, knowledge: "2025-01", release_date: "2025-09-30", last_updated: "2025-12-10", modalities: { input: ["text"], output: ["audio"] }, open_weights: false, limit: { context: 32768, output: 16384 } }, "google/gemini-embedding-2": { id: "google/gemini-embedding-2", name: "Gemini Embedding 2", description: "Multimodal embedding model mapping text, images, video, audio, and PDFs into a unified embedding space", family: "gemini", attachment: true, reasoning: false, tool_call: false, temperature: false, knowledge: "2025-11", release_date: "2026-04-22", last_updated: "2026-04-22", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 8192, output: 3072 } }, "google/gemma-4-E4B-it": { id: "google/gemma-4-E4B-it", name: "Gemma 4 E4B IT", description: "Open Gemma instruction model for efficient chat and self-hosted deployments", family: "gemma", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-02", last_updated: "2026-04-02", modalities: { input: ["text", "image", "audio"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/google/gemma-4-E4B-it" }] }, "google/gemma-4-E2B-it": { id: "google/gemma-4-E2B-it", name: "Gemma 4 E2B IT", description: "Open Gemma instruction model for efficient chat and self-hosted deployments", family: "gemma", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-02", last_updated: "2026-04-02", modalities: { input: ["text", "image", "audio"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/google/gemma-4-E2B-it" }] }, "google/gemini-3.6-flash": { id: "google/gemini-3.6-flash", name: "Gemini 3.6 Flash", description: "Fast Gemini model balancing multimodal reasoning, tool use, and cost", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-03", release_date: "2026-07-21", last_updated: "2026-07-21", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "SWE-Bench Pro", score: 58.7, metric: "resolve rate", harness: "Antigravity", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "DeepSWE", score: 49, metric: "resolve rate", variant: "high reasoning", version: "1.1", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "Terminal-Bench", score: 78, metric: "accuracy", harness: "Terminus 2", version: "2.1", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "MLE-Bench", score: 63.9, metric: "average position score", dataset: "Partial 30", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "GDPval-AA", score: 1421, metric: "Elo", version: "v2", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "OSWorld-Verified", score: 83, metric: "success rate", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "CharXiv Reasoning", score: 85.2, metric: "accuracy", variant: "no tools", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "CharXiv Reasoning", score: 89.4, metric: "accuracy", variant: "with tools", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "GDM-MRCR", score: 91.8, metric: "accuracy", variant: "128k average, 8-needle", version: "v2", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }, { name: "GDM-MRCR", score: 54, metric: "accuracy", variant: "1M pointwise, 8-needle", version: "v2", source: "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/", date: "2026-07-21" }] }, "google/gemini-3.1-flash-image": { id: "google/gemini-3.1-flash-image", name: "Nano Banana 2", description: "Image model for prompt-driven generation, editing, and visual design workflows", family: "gemini-flash", attachment: true, reasoning: true, tool_call: false, temperature: true, knowledge: "2025-01", release_date: "2026-05-28", last_updated: "2026-05-28", modalities: { input: ["text", "image", "video", "pdf"], output: ["text", "image"] }, open_weights: false, limit: { context: 131072, output: 32768 } }, "google/gemini-3.1-flash-image-preview": { id: "google/gemini-3.1-flash-image-preview", name: "Nano Banana 2", description: "Image model for prompt-driven generation, editing, and visual design workflows", family: "gemini-flash", attachment: true, reasoning: true, tool_call: false, temperature: true, knowledge: "2025-01", release_date: "2026-02-26", last_updated: "2026-02-26", modalities: { input: ["text", "image", "pdf"], output: ["text", "image"] }, open_weights: false, limit: { context: 65536, output: 65536 } }, "google/gemini-2.0-flash-lite": { id: "google/gemini-2.0-flash-lite", name: "Gemini 2.0 Flash-Lite", description: "Low-latency Gemini model for high-volume multimodal and agent workloads", family: "gemini-flash-lite", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-06", release_date: "2024-12-11", last_updated: "2024-12-11", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 8192 } }, "google/lyria-3-pro-preview": { id: "google/lyria-3-pro-preview", name: "Lyria 3 Pro Preview", description: "Music generation model for full-length songs from text or images with vocals and structure", family: "lyria", attachment: true, reasoning: false, tool_call: false, structured_output: false, temperature: true, release_date: "2026-03-25", last_updated: "2026-03-25", modalities: { input: ["text", "image"], output: ["text", "audio"] }, open_weights: false, limit: { context: 131072, output: 8192 } }, "google/gemini-3.1-flash-lite-preview": { id: "google/gemini-3.1-flash-lite-preview", name: "Gemini 3.1 Flash Lite Preview", description: "Low-latency Gemini model for high-volume multimodal and agent workloads", family: "gemini-flash-lite", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-03-03", last_updated: "2026-03-03", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/gemini-3-flash-preview": { id: "google/gemini-3-flash-preview", name: "Gemini 3 Flash Preview", description: "New Gemini flash lane bringing frontier-style multimodal reasoning to cheaper runs", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2025-12-17", last_updated: "2025-12-17", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "SWE-Bench Pro", score: 34.63, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }, { name: "SWE-Atlas Codebase QnA", score: 8.2, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 10, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 30.3, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }] }, "google/gemini-2.5-flash-image": { id: "google/gemini-2.5-flash-image", name: "Nano Banana", description: "Nano Banana image model for fast generation, edits, and character-consistent assets", family: "gemini-flash", attachment: true, reasoning: true, tool_call: false, temperature: true, knowledge: "2024-06", release_date: "2025-08-26", last_updated: "2025-08-26", modalities: { input: ["text", "image"], output: ["text", "image"] }, open_weights: false, limit: { context: 32768, output: 32768 } }, "google/gemini-2.5-pro": { id: "google/gemini-2.5-pro", name: "Gemini 2.5 Pro", description: "Google's proven reasoning model for coding, math, and multimodal analysis", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2025-06-17", last_updated: "2025-06-17", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "Aider Polyglot", score: 83.1, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-06-06" }, { name: "Artificial Analysis Coding Index", score: 32, metric: "index", source: "https://openrouter.ai/google/gemini-2.5-pro/benchmarks", date: "2026-06-02" }, { name: "SciCode", score: 42.8, metric: "percent correct", source: "https://openrouter.ai/google/gemini-2.5-pro/benchmarks", date: "2026-06-02" }, { name: "Terminal-Bench Hard", score: 26.5, metric: "success rate", source: "https://openrouter.ai/google/gemini-2.5-pro/benchmarks", date: "2026-06-02" }] }, "google/deep-research-max-preview-04-2026": { id: "google/deep-research-max-preview-04-2026", name: "Deep Research Max Preview", description: "Maximum-comprehensiveness agentic researcher for multi-step investigation, synthesis, and cited reports", family: "gemini-pro", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2025-01", release_date: "2026-04-21", last_updated: "2026-04-21", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text", "image"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "google/gemini-3.1-flash-live-preview": { id: "google/gemini-3.1-flash-live-preview", name: "Gemini 3.1 Flash Live Preview", description: "High-quality, low-latency Live API model for real-time dialogue and voice-first AI applications", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: false, temperature: true, knowledge: "2025-01", release_date: "2026-03-26", last_updated: "2026-03-26", modalities: { input: ["text", "image", "video", "audio"], output: ["text", "audio"] }, open_weights: false, limit: { context: 131072, output: 65536 } }, "google/gemini-3.7-flash": { id: "google/gemini-3.7-flash", name: "Gemini 3.7 Flash", description: "High-efficiency Gemini model for agentic workflows, coding, and multimodal reasoning", family: "gemini-flash", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-03", release_date: "2026-08-13", last_updated: "2026-08-13", modalities: { input: ["text", "image", "video", "audio", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 }, benchmarks: [{ name: "FrontierCode", score: 43.6, metric: "score", version: "1.1 Main", source: "https://deepmind.google/models/model-cards/gemini-3-7-flash/", date: "2026-08-13" }, { name: "DeepSWE", score: 65.3, metric: "resolve rate", version: "1.1", source: "https://deepmind.google/models/model-cards/gemini-3-7-flash/", date: "2026-08-13" }, { name: "Terminal-Bench", score: 85.8, metric: "accuracy", version: "2.1", source: "https://deepmind.google/models/model-cards/gemini-3-7-flash/", date: "2026-08-13" }, { name: "AutomationBench", score: 30.4, metric: "accuracy", dataset: "private set", source: "https://deepmind.google/models/model-cards/gemini-3-7-flash/", date: "2026-08-13" }, { name: "GDP.pdf", score: 34, metric: "accuracy", source: "https://deepmind.google/models/model-cards/gemini-3-7-flash/", date: "2026-08-13" }, { name: "GDM-MRCR", score: 97, metric: "accuracy", variant: "128k average, 8-needle", version: "v2", source: "https://deepmind.google/models/model-cards/gemini-3-7-flash/", date: "2026-08-13" }] }, "meta/llama-guard-3-8b": { id: "meta/llama-guard-3-8b", name: "Llama-Guard-3-8B", description: "Llama 3.1-based safety classifier for moderating prompts and model responses", family: "llama", attachment: false, reasoning: false, tool_call: false, temperature: true, knowledge: "2023-12", release_date: "2024-07-23", last_updated: "2024-07-23", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-Guard-3-8B" }] }, "meta/llama-3.3-70b-instruct": { id: "meta/llama-3.3-70b-instruct", name: "Llama-3.3-70B-Instruct", description: "Popular open Llama workhorse for multilingual chat, coding, and self-hosting", family: "llama", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2023-12", release_date: "2024-12-06", last_updated: "2024-12-06", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 10.7, metric: "index", source: "https://openrouter.ai/meta-llama/llama-3.3-70b-instruct/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 26, metric: "percent correct", source: "https://openrouter.ai/meta-llama/llama-3.3-70b-instruct/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 3, metric: "success rate", source: "https://openrouter.ai/meta-llama/llama-3.3-70b-instruct/benchmarks", date: "2026-03-11" }] }, "meta/llama-4-scout-17b-instruct": { id: "meta/llama-4-scout-17b-instruct", name: "Llama 4 Scout 17B Instruct", description: "Open Llama with long-context vision for efficient multimodal agents", family: "llama", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-08", release_date: "2025-04-05", last_updated: "2025-04-05", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 35e5, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct" }] }, "meta/muse-glimmer-30b": { id: "meta/muse-glimmer-30b", name: "Muse Glimmer 30B", description: "Muse Glimmer is a 30-billion-parameter open-weight multimodal model from Meta Superintelligence Labs, distilled from Muse Spark for always-on local agents, tool use, coding, and image understanding.", family: "muse", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-01-04", release_date: "2026-08-10", last_updated: "2026-08-10", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 }, license: "Apache 2.0", links: [{ label: "Announcement", url: "https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model", type: "announcement" }, { label: "Model card", url: "https://huggingface.co/meta-models/Muse-Glimmer-30B", type: "model_card" }, { label: "Developer docs", url: "https://developer.meta.com/ai/models/muse-glimmer/", type: "docs" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-models/Muse-Glimmer-30B" }], benchmarks: [{ name: "MCP Atlas", score: 75.5, metric: "success rate", variant: "public", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "DeepSearch QA", score: 74.6, source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "SWE-Bench Pro", score: 51.2, metric: "resolve rate", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "SWE-Bench Verified", score: 76, metric: "resolve rate", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "Terminal-Bench", score: 51.7, metric: "success rate", variant: "with terminus2", version: "2.1", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "OSWorld-Verified", score: 65.9, metric: "success rate", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "AIME 2026", score: 94.7, metric: "accuracy", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "GPQA Diamond", score: 83.5, metric: "accuracy", variant: "AA", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }, { name: "CharXiv Reasoning", score: 78.8, metric: "accuracy", source: "https://huggingface.co/meta-models/Muse-Glimmer-30B", date: "2026-08-10" }] }, "meta/llama-3.1-8b-instruct": { id: "meta/llama-3.1-8b-instruct", name: "Llama-3.1-8B-Instruct", description: "Compact open Llama model for lightweight chat, drafting, and self-hosting", family: "llama", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2023-12", release_date: "2024-07-23", last_updated: "2024-07-23", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct" }] }, "meta/muse-spark-1.2": { id: "meta/muse-spark-1.2", name: "Muse Spark 1.2", description: "Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1 with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows.", family: "muse", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-05", last_updated: "2026-08-05", modalities: { input: ["text", "image", "video", "pdf", "audio"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 131072 } }, "meta/muse-spark-1.1": { id: "meta/muse-spark-1.1", name: "Muse Spark 1.1", description: "Muse Spark is a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration.", family: "muse", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-08", last_updated: "2026-07-09", modalities: { input: ["text", "image", "pdf", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 32e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 61.5, metric: "resolve rate", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "Terminal-Bench", score: 80, metric: "success rate", version: "2.1", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "DeepSWE", score: 53.3, metric: "resolve rate", version: "1.1", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "MCP Atlas", score: 88.1, metric: "success rate", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "JobBench", score: 54.7, metric: "success rate", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "Toolathlon-Verified", score: 75.6, metric: "success rate", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "Humanity's Last Exam", score: 62.1, metric: "accuracy", variant: "with tools", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "OSWorld-Verified", score: 80.8, metric: "success rate", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "Finance Agent", score: 57.2, metric: "accuracy", version: "v2", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "CharXiv Reasoning", score: 88.4, metric: "accuracy", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }, { name: "BabyVision", score: 76.3, metric: "accuracy", source: "https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/", date: "2026-07-09" }] }, "meta/llama-3.2-1b": { id: "meta/llama-3.2-1b", name: "Llama-3.2-1B", description: "Compact open Llama base model for lightweight and on-device use", family: "llama", attachment: false, reasoning: false, tool_call: false, temperature: true, knowledge: "2023-12", release_date: "2024-09-25", last_updated: "2024-09-25", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, license: "Llama 3.2 Community License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-3.2-1B" }] }, "meta/llama-4-maverick-17b-instruct": { id: "meta/llama-4-maverick-17b-instruct", name: "Llama 4 Maverick 17B Instruct", description: "Open multimodal Llama for strong reasoning with efficient everyday serving", family: "llama", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-08", release_date: "2025-04-05", last_updated: "2025-04-05", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct" }], benchmarks: [{ name: "Aider Polyglot", score: 15.6, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-04-06" }, { name: "SWE-Bench Pro", score: 5.24, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "meta/llama-3.2-3b": { id: "meta/llama-3.2-3b", name: "Llama-3.2-3B", description: "Small open Llama base model for lightweight text generation and self-hosting", family: "llama", attachment: false, reasoning: false, tool_call: false, temperature: true, knowledge: "2023-12", release_date: "2024-09-25", last_updated: "2024-09-25", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, license: "Llama 3.2 Community License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-3.2-3B" }] }, "meta/llama-3.2-11b-vision-instruct": { id: "meta/llama-3.2-11b-vision-instruct", name: "Llama-3.2-11B-Vision-Instruct", description: "Open multimodal Llama model for image understanding, captioning, and visual QA", family: "llama", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2023-12", release_date: "2024-09-25", last_updated: "2024-09-25", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4096 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct" }] }, "sdaia/allam-2-7b": { id: "sdaia/allam-2-7b", name: "ALLaM-2-7b", description: "ALLaM-2-7b instruction tuned model by SDAIA", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2025-01-23", last_updated: "2025-01-23", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 4096, output: 4096 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/ALLaM-AI/ALLaM-2.0-7B-Instruct" }] }, "poolside/laguna-m.1": { id: "poolside/laguna-m.1", name: "Laguna M.1", description: "Poolside's open-weight model for agentic coding and long-horizon work", family: "laguna", attachment: false, reasoning: true, tool_call: true, structured_output: false, temperature: true, release_date: "2026-04-28", last_updated: "2026-06-13", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 32768 } }, "poolside/laguna-s-2.1": { id: "poolside/laguna-s-2.1", name: "Laguna S 2.1", description: "Agentic coding model from Poolside in the XS size class for local deployment", family: "laguna", attachment: false, reasoning: true, tool_call: true, structured_output: false, temperature: true, release_date: "2026-07-21", last_updated: "2026-07-21", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 32768 } }, "poolside/laguna-xs-2.1": { id: "poolside/laguna-xs-2.1", name: "Laguna XS 2.1", description: "Agentic coding model from Poolside in the XS size class for local deployment", family: "laguna", attachment: false, reasoning: true, tool_call: true, structured_output: false, temperature: true, release_date: "2026-07-02", last_updated: "2026-07-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 32768 }, benchmarks: [{ name: "SWE-Bench Verified", score: 70.9, metric: "resolved", harness: "Harbor", source: "https://poolside.ai/blog/introducing-laguna-xs-2-1", date: "2026-07-02" }, { name: "SWE-Bench Multilingual", score: 63.1, metric: "resolve rate", harness: "Harbor", source: "https://poolside.ai/blog/introducing-laguna-xs-2-1", date: "2026-07-02" }, { name: "SWE-Bench Pro", score: 47.6, metric: "resolve rate", harness: "Harbor", source: "https://poolside.ai/blog/introducing-laguna-xs-2-1", date: "2026-07-02" }, { name: "Terminal-Bench", score: 37.5, metric: "success rate", harness: "Harbor", version: "2.0", source: "https://poolside.ai/blog/introducing-laguna-xs-2-1", date: "2026-07-02" }] }, "poolside/laguna-xs.2": { id: "poolside/laguna-xs.2", name: "Laguna XS.2", description: "Agentic coding model from Poolside in the XS size class for local deployment", family: "laguna", attachment: false, reasoning: true, tool_call: true, structured_output: false, temperature: true, release_date: "2026-04-28", last_updated: "2026-06-13", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 32768 } }, "bytedance-seed/seed-2.0-code": { id: "bytedance-seed/seed-2.0-code", name: "Seed 2.0 Code", description: "ByteDance Seed coding model for multimodal software engineering and long-running agents", family: "seed", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-14", last_updated: "2026-02-14", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 262144, output: 131072 } }, "bytedance-seed/seed-2.1-turbo": { id: "bytedance-seed/seed-2.1-turbo", name: "Seed 2.1 Turbo", description: "Faster ByteDance Seed 2.1 model for multimodal reasoning and latency-sensitive agent workflows", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-23", last_updated: "2026-06-23", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 256e3 } }, "bytedance-seed/seed-1-6": { id: "bytedance-seed/seed-1-6", name: "Seed 1.6", description: "ByteDance Seed model for long-context reasoning, instruction following, and tool-assisted tasks", family: "seed", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-10-15", last_updated: "2025-10-15", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 64e3 } }, "bytedance-seed/seed-character": { id: "bytedance-seed/seed-character", name: "Seed Character", description: "ByteDance Seed model optimized for character-driven dialogue and consistent conversational behavior", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-23", last_updated: "2026-06-23", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 256e3 } }, "bytedance-seed/seed-1-6-vision": { id: "bytedance-seed/seed-1-6-vision", name: "Seed 1.6 Vision", description: "ByteDance Seed multimodal model for image understanding, visual reasoning, and tool-assisted tasks", family: "seed", attachment: true, reasoning: false, tool_call: true, temperature: true, release_date: "2025-08-15", last_updated: "2025-08-15", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 32e3 } }, "bytedance-seed/seed-2.0-pro": { id: "bytedance-seed/seed-2.0-pro", name: "Seed 2.0 Pro", description: "Flagship ByteDance Seed 2.0 model for complex multimodal reasoning and long-horizon agent workflows", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-14", last_updated: "2026-02-14", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 128e3 } }, "bytedance-seed/seed-evolving": { id: "bytedance-seed/seed-evolving", name: "Seed Evolving", description: "Rolling ByteDance Seed model for rapidly updated reasoning, coding, and agent capabilities", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-23", last_updated: "2026-06-23", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 256e3 } }, "bytedance-seed/seed-2.1-pro": { id: "bytedance-seed/seed-2.1-pro", name: "Seed 2.1 Pro", description: "Flagship ByteDance Seed 2.1 model for complex multimodal reasoning, coding, and agents", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-23", last_updated: "2026-06-23", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 256e3 } }, "bytedance-seed/seed-1-6-flash": { id: "bytedance-seed/seed-1-6-flash", name: "Seed 1.6 Flash", description: "Low-latency ByteDance Seed model for high-throughput chat, extraction, and lightweight tool use", family: "seed", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2025-08-28", last_updated: "2025-08-28", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 32e3 } }, "bytedance-seed/seed-2.0-mini": { id: "bytedance-seed/seed-2.0-mini", name: "Seed 2.0 Mini", description: "Lightweight ByteDance Seed 2.0 model for low-latency multimodal reasoning and high-volume tasks", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-14", last_updated: "2026-02-14", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 32e3 } }, "bytedance-seed/seed-1-8": { id: "bytedance-seed/seed-1-8", name: "Seed 1.8", description: "ByteDance Seed model for multimodal reasoning, long-context analysis, and agent workflows", family: "seed", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-12-28", last_updated: "2025-12-28", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 64e3 } }, "bytedance-seed/seed-2.0-lite": { id: "bytedance-seed/seed-2.0-lite", name: "Seed 2.0 Lite", description: "Cost-efficient ByteDance Seed 2.0 model for production chat, analysis, and structured generation", family: "seed", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-14", last_updated: "2026-02-14", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 32e3 } }, "moonshotai/kimi-k2.7-code": { id: "moonshotai/kimi-k2.7-code", name: "Kimi K2.7 Code", description: "Coding-focused Kimi model, stronger on long-horizon repo work with less overthinking", family: "kimi-k2", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-01", release_date: "2026-06-12", last_updated: "2026-06-12", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/moonshotai/Kimi-K2.7-Code" }], benchmarks: [{ name: "Kimi Code Bench", score: 62, harness: "Kimi Code CLI", version: "v2", source: "https://huggingface.co/moonshotai/Kimi-K2.7-Code", date: "2026-06-12" }, { name: "Program Bench", score: 53.6, harness: "Kimi Code CLI", source: "https://huggingface.co/moonshotai/Kimi-K2.7-Code", date: "2026-06-12" }, { name: "MLS Bench Lite", score: 35.1, harness: "Kimi Code CLI", source: "https://huggingface.co/moonshotai/Kimi-K2.7-Code", date: "2026-06-12" }, { name: "MCP Atlas", score: 76, metric: "success rate", harness: "Kimi Code CLI", source: "https://huggingface.co/moonshotai/Kimi-K2.7-Code", date: "2026-06-12" }, { name: "MCP Mark Verified", score: 81.1, metric: "success rate", harness: "Kimi Code CLI", source: "https://huggingface.co/moonshotai/Kimi-K2.7-Code", date: "2026-06-12" }, { name: "Kimi Claw 24/7 Bench", score: 46.9, harness: "Kimi Code CLI", source: "https://huggingface.co/moonshotai/Kimi-K2.7-Code", date: "2026-06-12" }] }, "moonshotai/kimi-k3": { id: "moonshotai/kimi-k3", name: "Kimi K3", description: "Multimodal Kimi model with 1M context and toggleable max-effort thinking for long-horizon agent work", family: "kimi-k3", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, release_date: "2026-07-16", last_updated: "2026-07-16", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 131072 }, benchmarks: [{ name: "DeepSWE", score: 67.5, metric: "resolve rate", harness: "Kimi Code", variant: "max effort", version: "1.1", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "Terminal-Bench", score: 88.3, metric: "accuracy", harness: "Kimi Code", variant: "max effort", version: "2.1", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "FrontierSWE", score: 81.2, metric: "dominance score", harness: "Kimi Code", variant: "max effort", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "Program Bench", score: 77.8, metric: "score", harness: "Kimi Code", variant: "max effort", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "SWE Marathon", score: 42, metric: "resolve rate", harness: "Claude Code", variant: "max effort", version: "1.1", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "GDPval-AA", score: 1668, metric: "Elo", variant: "max effort", version: "v2", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "AA-Briefcase", score: 1548, metric: "Elo", variant: "max effort", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "AutomationBench", score: 30.8, metric: "success rate", variant: "max effort", dataset: "600-task public subset", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "JobBench", score: 52.9, metric: "score", variant: "max effort", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "SpreadsheetBench", score: 34.8, metric: "score", harness: "Claude Code", variant: "max effort", version: "2", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "BrowseComp", score: 91.2, metric: "accuracy", variant: "max effort, context compaction", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "CharXiv Reasoning", score: 91.3, metric: "accuracy", variant: "max effort, with tools", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }, { name: "ZeroBench", score: 41, metric: "pass@5", variant: "max effort, with tools", source: "https://www.kimi.com/blog/kimi-k3", date: "2026-07-16" }] }, "moonshotai/kimi-k2-thinking-turbo": { id: "moonshotai/kimi-k2-thinking-turbo", name: "Kimi K2 Thinking Turbo", description: "Kimi reasoning model for long-horizon research, planning, and tool use", family: "kimi-thinking", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-08", release_date: "2025-11-06", last_updated: "2025-11-06", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/moonshotai/Kimi-K2-Thinking" }] }, "moonshotai/kimi-k2.5": { id: "moonshotai/kimi-k2.5", name: "Kimi K2.5", description: "Earlier Kimi frontier model for long-context agents, coding, and multimodal work", family: "kimi-k2", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-01", release_date: "2026-01", last_updated: "2026-01", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/moonshotai/Kimi-K2.5" }], benchmarks: [{ name: "SWE-Bench Verified", score: 70.8, metric: "resolved", source: "https://www.swebench.com/" }, { name: "SWE-Atlas Codebase QnA", score: 13.1, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 20.95, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 25.77, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }] }, "moonshotai/kimi-k2.7-code-highspeed": { id: "moonshotai/kimi-k2.7-code-highspeed", name: "Kimi K2.7 Code Highspeed", description: "Lower-latency Kimi Code variant for interactive edits and coding-agent loops", family: "kimi-k2", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-01", release_date: "2026-06-12", last_updated: "2026-06-12", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/moonshotai/Kimi-K2.7-Code" }] }, "moonshotai/kimi-k2-thinking": { id: "moonshotai/kimi-k2-thinking", name: "Kimi K2 Thinking", description: "Thinking Kimi model for slower research passes, planning, and hard technical questions", family: "kimi-thinking", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-08", release_date: "2025-11-06", last_updated: "2025-11-06", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/moonshotai/Kimi-K2-Thinking" }], benchmarks: [{ name: "SWE-Bench Verified", score: 71.3, metric: "resolved", source: "https://huggingface.co/moonshotai/Kimi-K2-Thinking" }] }, "moonshotai/kimi-k2.6": { id: "moonshotai/kimi-k2.6", name: "Kimi K2.6", description: "Multimodal Kimi workhorse for agent loops, coding tasks, and visual context", family: "kimi-k2", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-01", release_date: "2026-04-21", last_updated: "2026-04-21", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/moonshotai/Kimi-K2.6" }], benchmarks: [{ name: "SWE-Bench Verified", score: 80.2, metric: "resolved", source: "https://huggingface.co/moonshotai/Kimi-K2.6" }, { name: "Artificial Analysis Coding Agent Index", score: 50.5, metric: "average pass@1", harness: "Claude Code", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 59.8, metric: "pass@1", harness: "Claude Code", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 27.3, metric: "pass@1", harness: "Claude Code", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 64.3, metric: "pass@1", harness: "Claude Code", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }] }, "anthropic/claude-opus-4-20250514": { id: "anthropic/claude-opus-4-20250514", name: "Claude Opus 4", description: "Flagship Claude model for deep reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-05-22", last_updated: "2025-05-22", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 32e3 }, benchmarks: [{ name: "Aider Polyglot", score: 72, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-25" }] }, "anthropic/claude-opus-4-7": { id: "anthropic/claude-opus-4-7", name: "Claude Opus 4.7", description: "Stronger Opus tier for advanced software work and high-stakes reasoning", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2026-01-31", release_date: "2026-04-16", last_updated: "2026-04-16", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 64.3, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "Terminal-Bench", score: 66.1, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "SWE-Atlas Refactoring", score: 48.57, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "Artificial Analysis Coding Agent Index", score: 66.6, metric: "average pass@1", harness: "Claude Code", variant: "max", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 81, metric: "pass@1", harness: "Claude Code", variant: "max", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 44.9, metric: "pass@1", harness: "Claude Code", variant: "max", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 73.8, metric: "pass@1", harness: "Claude Code", variant: "max", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Artificial Analysis Coding Agent Index", score: 61.2, metric: "average pass@1", harness: "Cursor CLI", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 78.4, metric: "pass@1", harness: "Cursor CLI", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 34.4, metric: "pass@1", harness: "Cursor CLI", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 70.6, metric: "pass@1", harness: "Cursor CLI", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Artificial Analysis Coding Agent Index", score: 59.9, metric: "average pass@1", harness: "Claude Code", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 71.7, metric: "pass@1", harness: "Claude Code", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 36.4, metric: "pass@1", harness: "Claude Code", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 71.4, metric: "pass@1", harness: "Claude Code", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "GPQA Diamond", score: 94.2, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 46.9, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 54.7, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "OSWorld-Verified", score: 78, metric: "success rate", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }] }, "anthropic/claude-mythos-5": { id: "anthropic/claude-mythos-5", name: "Claude Mythos 5", description: "Restricted Claude model for advanced cybersecurity and biology research workflows", family: "claude-mythos", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2026-01-31", release_date: "2026-06-09", last_updated: "2026-06-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 } }, "anthropic/claude-opus-4-1-20250805": { id: "anthropic/claude-opus-4-1-20250805", name: "Claude Opus 4.1", description: "Flagship Claude model for deep reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-08-05", last_updated: "2025-08-05", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 32e3 } }, "anthropic/claude-3-5-sonnet-20241022": { id: "anthropic/claude-3-5-sonnet-20241022", name: "Claude Sonnet 3.5 v2", description: "Balanced Claude model for coding, analysis, agent workflows, and cost control", family: "claude-sonnet", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-04-30", release_date: "2024-10-22", last_updated: "2024-10-22", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 8192 }, benchmarks: [{ name: "Aider Polyglot", score: 51.6, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-01-17" }] }, "anthropic/claude-opus-4-8": { id: "anthropic/claude-opus-4-8", name: "Claude Opus 4.8", description: "Top Claude Opus tier for the hardest reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2026-01", release_date: "2026-05-28", last_updated: "2026-05-28", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 69.2, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "Terminal-Bench", score: 74.6, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "SWE-Bench Verified", score: 88.6, metric: "resolved", source: "https://benchlm.ai/benchmarks/sweVerified" }, { name: "Humanity's Last Exam", score: 49.8, metric: "accuracy", variant: "no tools", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "Humanity's Last Exam", score: 57.9, metric: "accuracy", variant: "with tools", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "OSWorld-Verified", score: 83.4, metric: "success rate", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "FrontierCode", score: 13.4, metric: "pass rate", variant: "high effort", dataset: "Diamond", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }] }, "anthropic/claude-3-5-haiku-20241022": { id: "anthropic/claude-3-5-haiku-20241022", name: "Claude Haiku 3.5", description: "Fast Claude model for responsive assistance, classification, and lightweight agents", family: "claude-haiku", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-07-31", release_date: "2024-10-22", last_updated: "2024-10-22", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 8192 }, benchmarks: [{ name: "Aider Polyglot", score: 28, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2024-12-21" }] }, "anthropic/claude-opus-4-1": { id: "anthropic/claude-opus-4-1", name: "Claude Opus 4.1 (latest)", description: "Flagship Claude model for deep reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-08-05", last_updated: "2025-08-05", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 32e3 } }, "anthropic/claude-sonnet-5": { id: "anthropic/claude-sonnet-5", name: "Claude Sonnet 5", description: "Everyday Claude agent model for coding, planning, browsing, and general work", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2026-01-31", release_date: "2026-06-30", last_updated: "2026-06-30", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Verified", score: 85.2, metric: "resolved", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "SWE-Bench Pro", score: 63.2, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "SWE-Bench Multilingual", score: 78.3, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "Terminal-Bench", score: 80.4, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "OSWorld-Verified", score: 81.2, metric: "success rate", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "BrowseComp", score: 84.7, metric: "accuracy", variant: "single agent", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "FrontierCode", score: 38.8, metric: "pass rate", version: "v1", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }] }, "anthropic/claude-3-7-sonnet-20250219": { id: "anthropic/claude-3-7-sonnet-20250219", name: "Claude Sonnet 3.7", description: "Balanced Claude model for coding, analysis, agent workflows, and cost control", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-10-31", release_date: "2025-02-19", last_updated: "2025-02-19", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 }, benchmarks: [{ name: "Aider Polyglot", score: 64.9, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-02-24" }] }, "anthropic/claude-sonnet-4-5-20250929": { id: "anthropic/claude-sonnet-4-5-20250929", name: "Claude Sonnet 4.5", description: "Balanced Claude model for coding, analysis, agent workflows, and cost control", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-07-31", release_date: "2025-09-29", last_updated: "2025-09-29", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 } }, "anthropic/claude-opus-4-5": { id: "anthropic/claude-opus-4-5", name: "Claude Opus 4.5 (latest)", description: "Flagship Claude model for deep reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-05", release_date: "2025-11-24", last_updated: "2025-11-24", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 } }, "anthropic/claude-sonnet-4-6": { id: "anthropic/claude-sonnet-4-6", name: "Claude Sonnet 4.6", description: "Claude workhorse for coding agents, careful analysis, and production cost control", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-08-31", release_date: "2026-02-17", last_updated: "2026-03-13", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 64e3 }, benchmarks: [{ name: "SWE-Atlas Codebase QnA", score: 31.2, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 32.21, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 31.76, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "Artificial Analysis Coding Agent Index", score: 49.4, metric: "average pass@1", harness: "Claude Code", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 70.3, metric: "pass@1", harness: "Claude Code", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 14.9, metric: "pass@1", harness: "Claude Code", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 63.1, metric: "pass@1", harness: "Claude Code", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 67, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "Humanity's Last Exam", score: 34.6, metric: "accuracy", variant: "no tools", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "Humanity's Last Exam", score: 46.8, metric: "accuracy", variant: "with tools", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }, { name: "OSWorld-Verified", score: 78.5, metric: "success rate", source: "https://www.anthropic.com/news/claude-sonnet-5", date: "2026-06-30" }] }, "anthropic/claude-opus-4-0": { id: "anthropic/claude-opus-4-0", name: "Claude Opus 4 (latest)", description: "Flagship Claude model for deep reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-05-22", last_updated: "2025-05-22", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 32e3 }, benchmarks: [{ name: "Aider Polyglot", score: 72, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-25" }] }, "anthropic/claude-haiku-4-5-20251001": { id: "anthropic/claude-haiku-4-5-20251001", name: "Claude Haiku 4.5", description: "Fast Claude model for responsive assistance, classification, and lightweight agents", family: "claude-haiku", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-02-28", release_date: "2025-10-15", last_updated: "2025-10-15", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 } }, "anthropic/claude-sonnet-4-5": { id: "anthropic/claude-sonnet-4-5", name: "Claude Sonnet 4.5 (latest)", description: "Balanced Claude model for coding, analysis, agent workflows, and cost control", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-07-31", release_date: "2025-09-29", last_updated: "2025-09-29", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 43.6, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "anthropic/claude-opus-4-5-20251101": { id: "anthropic/claude-opus-4-5-20251101", name: "Claude Opus 4.5", description: "Flagship Claude model for deep reasoning, coding, and long-horizon agents", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-05", release_date: "2025-11-01", last_updated: "2025-11-01", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 45.89, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "anthropic/claude-fable-5": { id: "anthropic/claude-fable-5", name: "Claude Fable 5", description: "Claude model for creative writing, analysis, and controlled agent workflows", family: "claude-fable", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2026-01-31", release_date: "2026-06-09", last_updated: "2026-06-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 80.3, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "SWE-Bench Verified", score: 95, metric: "resolved", source: "https://benchlm.ai/benchmarks/sweVerified" }, { name: "Terminal-Bench", score: 88, metric: "success rate", version: "2.1", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "Humanity's Last Exam", score: 59, metric: "accuracy", variant: "no tools", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "Humanity's Last Exam", score: 64.5, metric: "accuracy", variant: "with tools", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "OSWorld-Verified", score: 85, metric: "success rate", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "FrontierCode", score: 29.3, metric: "pass rate", variant: "high effort", dataset: "Diamond", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "GDPval-AA", score: 1932, metric: "Elo", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }, { name: "AutomationBench", score: 17.4, metric: "success rate", source: "https://www.anthropic.com/news/claude-fable-5-mythos-5", date: "2026-06-09" }] }, "anthropic/claude-sonnet-4-20250514": { id: "anthropic/claude-sonnet-4-20250514", name: "Claude Sonnet 4", description: "Balanced Claude model for coding, analysis, agent workflows, and cost control", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-05-22", last_updated: "2025-05-22", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 }, benchmarks: [{ name: "Aider Polyglot", score: 61.3, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-24" }] }, "anthropic/claude-sonnet-4-0": { id: "anthropic/claude-sonnet-4-0", name: "Claude Sonnet 4 (latest)", description: "Balanced Claude model for coding, analysis, agent workflows, and cost control", family: "claude-sonnet", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-05-22", last_updated: "2025-05-22", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 }, benchmarks: [{ name: "Aider Polyglot", score: 61.3, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-24" }, { name: "SWE-Bench Pro", score: 42.7, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "anthropic/claude-haiku-4-5": { id: "anthropic/claude-haiku-4-5", name: "Claude Haiku 4.5 (latest)", description: "Fast Claude lane for lightweight agents, office tasks, and responsive chat", family: "claude-haiku", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-02-28", release_date: "2025-10-15", last_updated: "2025-10-15", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 64e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 39.45, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "anthropic/claude-opus-4-6": { id: "anthropic/claude-opus-4-6", name: "Claude Opus 4.6", description: "High-end Claude for difficult coding, planning, and slower expert reasoning", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-05-31", release_date: "2026-02-05", last_updated: "2026-03-13", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 51.9, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }, { name: "SWE-Atlas Codebase QnA", score: 33.3, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Codebase QnA", score: 30, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 35.58, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 36.67, metric: "score", harness: "Claude Code", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "SWE-Atlas Test Writing", score: 36.08, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "Artificial Analysis Coding Agent Index", score: 51.3, metric: "average pass@1", harness: "Claude Code", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 71.9, metric: "pass@1", harness: "Claude Code", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 11.8, metric: "pass@1", harness: "Claude Code", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 70.2, metric: "pass@1", harness: "Claude Code", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }] }, "anthropic/claude-3-haiku-20240307": { id: "anthropic/claude-3-haiku-20240307", name: "Claude Haiku 3", description: "Legacy model retained for compatibility with older integrations", family: "claude-haiku", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2023-08-31", release_date: "2024-03-13", last_updated: "2024-03-13", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 4096 } }, "anthropic/claude-opus-5": { id: "anthropic/claude-opus-5", name: "Claude Opus 5", description: "Strongest Claude Opus model for coding, agents, and professional work", family: "claude-opus", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2026-05", release_date: "2026-07-24", last_updated: "2026-07-24", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Verified", score: 96, metric: "resolved", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "SWE-Bench Pro", score: 79.2, metric: "resolve rate", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "SWE-Bench Multilingual", score: 89.5, metric: "resolve rate", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "SWE-Bench Multimodal", score: 59.4, metric: "resolve rate", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "DeepSWE", score: 68.8, metric: "resolve rate", version: "1.1", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "FrontierCode", score: 53.4, metric: "mean@5", variant: "medium effort", dataset: "Main", version: "1.1", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "Frontier-Bench", score: 43.3, metric: "mean reward", harness: "mini-SWE-agent", variant: "max effort", version: "v0.1", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "BrowseComp", score: 90.8, metric: "accuracy", variant: "single agent", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "Humanity's Last Exam", score: 56.3, metric: "accuracy", variant: "no tools", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "Humanity's Last Exam", score: 64.7, metric: "accuracy", variant: "with tools", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "DeepSearchQA", score: 95, metric: "F1", variant: "max effort", source: "https://www.anthropic.com/news/claude-opus-5", date: "2026-07-24" }, { name: "OSWorld", score: 70.6, metric: "success rate", version: "2.0", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "GDPval-AA", score: 1861, metric: "Elo", variant: "max effort", version: "v2", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "AA-Briefcase", score: 1720, metric: "Elo", variant: "max effort", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "AutomationBench", score: 26, metric: "success rate", variant: "max effort", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "ARC-AGI-1", score: 97.5, metric: "accuracy", variant: "max effort", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "ARC-AGI-2", score: 90.4, metric: "accuracy", variant: "max effort", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "ARC-AGI-3", score: 30.2, metric: "RHAE", variant: "high effort", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }, { name: "HealthBench Professional", score: 59.8, metric: "score", variant: "max effort", source: "https://www.anthropic.com/claude-opus-5-system-card", date: "2026-07-24" }] }, "cohere/c4ai-aya-expanse-32b": { id: "cohere/c4ai-aya-expanse-32b", name: "Aya Expanse 32B", description: "Open multilingual model optimized for generation across 23 languages", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2024-10-24", last_updated: "2024-10-24", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4e3 }, license: "CC-BY-NC-4.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/aya-expanse-32b" }] }, "cohere/command-a-reasoning-08-2025": { id: "cohere/command-a-reasoning-08-2025", name: "Command A Reasoning", description: "Cohere reasoning model for multilingual enterprise agents, tools, and complex workflows", family: "command-a", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2025-08-21", last_updated: "2025-08-21", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 32e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-a-reasoning-08-2025" }] }, "cohere/c4ai-aya-vision-32b": { id: "cohere/c4ai-aya-vision-32b", name: "Aya Vision 32B", description: "Open multilingual vision model for OCR, visual reasoning, and image question answering", attachment: true, reasoning: false, tool_call: false, temperature: true, release_date: "2025-03-04", last_updated: "2025-05-14", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 16e3, output: 4e3 }, license: "CC-BY-NC-4.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/aya-vision-32b" }] }, "cohere/command-r-plus-08-2024": { id: "cohere/command-r-plus-08-2024", name: "Command R+", description: "Cohere's RAG workhorse for long-context enterprise search and tool use", family: "command-r", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2024-08-30", last_updated: "2024-08-30", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-r-plus-08-2024" }] }, "cohere/command-a-translate-08-2025": { id: "cohere/command-a-translate-08-2025", name: "Command A Translate", description: "Translation model for multilingual conversion, localization, and cross-language workflows", family: "command-a", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2025-08-28", last_updated: "2025-08-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 8e3, output: 8e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-a-translate-08-2025" }] }, "cohere/command-r7b-arabic-02-2025": { id: "cohere/command-r7b-arabic-02-2025", name: "Command R7B Arabic", description: "Open Command R model optimized for Arabic enterprise chat, RAG, and cultural knowledge", family: "command-r", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2025-02-27", last_updated: "2025-02-27", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-r7b-arabic-02-2025" }] }, "cohere/c4ai-aya-expanse-8b": { id: "cohere/c4ai-aya-expanse-8b", name: "Aya Expanse 8B", description: "Compact open multilingual model optimized for generation across 23 languages", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2024-10-24", last_updated: "2024-10-24", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 8e3, output: 4e3 }, license: "CC-BY-NC-4.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/aya-expanse-8b" }] }, "cohere/command-a-vision-07-2025": { id: "cohere/command-a-vision-07-2025", name: "Command A Vision", description: "Cohere vision model for multilingual document analysis, OCR, and image understanding", family: "command-a", attachment: true, reasoning: false, tool_call: false, temperature: true, knowledge: "2024-06-01", release_date: "2025-07-31", last_updated: "2025-07-31", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 8e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-a-vision-07-2025" }] }, "cohere/command-a-plus-05-2026": { id: "cohere/command-a-plus-05-2026", name: "Command A Plus", description: "Cohere's stronger command model for multilingual agents and enterprise workflows", family: "command-a", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-04-01", release_date: "2026-05-20", last_updated: "2026-06-09", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 64e3 } }, "cohere/command-a-03-2025": { id: "cohere/command-a-03-2025", name: "Command A", description: "Cohere command model for multilingual enterprise agents, tools, and chat", family: "command-a", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2025-03-13", last_updated: "2025-03-13", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 8e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-a-03-2025" }], benchmarks: [{ name: "Aider Polyglot", score: 12, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-03-14" }] }, "cohere/command-r7b-12-2024": { id: "cohere/command-r7b-12-2024", name: "Command R7B", description: "Cohere retrieval model for long-context chat and enterprise RAG workflows", family: "command-r", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2024-12-02", last_updated: "2024-12-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-r7b-12-2024" }] }, "cohere/command-r-08-2024": { id: "cohere/command-r-08-2024", name: "Command R", description: "Cohere retrieval model for long-context chat and enterprise RAG workflows", family: "command-r", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-06-01", release_date: "2024-08-30", last_updated: "2024-08-30", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 4e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/c4ai-command-r-08-2024" }] }, "cohere/north-mini-code-1-0": { id: "cohere/north-mini-code-1-0", name: "North Mini Code", description: "Cohere coding model for practical software engineering and agentic edits", family: "north", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-09-23", release_date: "2026-06-09", last_updated: "2026-06-09", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 64e3 }, benchmarks: [{ name: "SWE-Bench Verified", score: 67.6, metric: "resolved", harness: "SWE-agent", source: "https://huggingface.co/CohereLabs/North-Mini-Code-1.0", date: "2026-06-09" }, { name: "SWE-Bench Pro", score: 40.2, metric: "resolve rate", harness: "SWE-agent", source: "https://huggingface.co/CohereLabs/North-Mini-Code-1.0", date: "2026-06-09" }, { name: "Artificial Analysis Intelligence Index", score: 27.6, metric: "index score", source: "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model", date: "2026-06-09" }, { name: "Artificial Analysis Coding Index", score: 33.4, metric: "index score", source: "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model", date: "2026-06-09" }, { name: "GDPval-AA", score: 14, metric: "win rate", source: "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model", date: "2026-06-09" }, { name: "\u03C4\xB2-Bench Telecom", score: 37, metric: "success rate", source: "https://artificialanalysis.ai/articles/north-mini-code-cohere-s-small-coding-focused-moe-model", date: "2026-06-09" }] }, "cohere/c4ai-aya-vision-8b": { id: "cohere/c4ai-aya-vision-8b", name: "Aya Vision 8B", description: "Compact open multilingual vision model for OCR and visual question answering", attachment: true, reasoning: false, tool_call: false, temperature: true, release_date: "2025-03-04", last_updated: "2025-05-14", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 16e3, output: 4e3 }, license: "CC-BY-NC-4.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/CohereLabs/aya-vision-8b" }] }, "openai/gpt-4o": { id: "openai/gpt-4o", name: "GPT-4o", description: "Omni-era GPT for multimodal chat, practical coding, and general assistants", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2023-09", release_date: "2024-05-13", last_updated: "2024-08-06", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 }, benchmarks: [{ name: "Aider Polyglot", score: 23.1, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2024-12-30" }] }, "openai/gpt-image-1.5": { id: "openai/gpt-image-1.5", name: "GPT-Image-1.5", description: "Image model for prompt-driven generation, editing, and visual design workflows", family: "gpt-image", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2025-11-25", last_updated: "2025-11-25", modalities: { input: ["text", "image"], output: ["text", "image"] }, open_weights: false, limit: { context: 0, output: 0 } }, "openai/gpt-5.3-chat-latest": { id: "openai/gpt-5.3-chat-latest", name: "GPT-5.3 Chat (latest)", description: "Chat-tuned GPT model for conversational assistance, writing, and tool workflows", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-08-31", release_date: "2026-03-03", last_updated: "2026-03-03", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 } }, "openai/gpt-5-nano": { id: "openai/gpt-5-nano", name: "GPT-5 Nano", description: "Tiny GPT-5 lane for routing, extraction, classification, and bulk jobs", family: "gpt-nano", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-05-30", release_date: "2025-08-07", last_updated: "2025-08-07", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-4o-mini": { id: "openai/gpt-4o-mini", name: "GPT-4o mini", description: "Small omni GPT for cheap multimodal assistance and production-scale traffic", family: "gpt-mini", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2023-09", release_date: "2024-07-18", last_updated: "2024-07-18", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 }, benchmarks: [{ name: "Aider Polyglot", score: 3.6, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2024-12-21" }, { name: "SciCode", score: 22.9, metric: "percent correct", source: "https://openrouter.ai/openai/gpt-4o-mini/benchmarks", date: "2026-03-11" }] }, "openai/gpt-3.5-turbo": { id: "openai/gpt-3.5-turbo", name: "GPT-3.5-turbo", description: "Compact GPT model for low-latency assistance and high-volume workloads", family: "gpt", attachment: false, reasoning: false, tool_call: false, structured_output: false, temperature: true, knowledge: "2021-09-01", release_date: "2023-03-01", last_updated: "2023-11-06", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 16385, output: 4096 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 10.7, metric: "index", source: "https://openrouter.ai/openai/gpt-3.5-turbo/benchmarks", date: "2026-03-11" }] }, "openai/o1-pro": { id: "openai/o1-pro", name: "o1-pro", description: "O-series reasoning model for hard analysis, math, coding, and planning", family: "o-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2023-09", release_date: "2025-03-19", last_updated: "2025-03-19", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 } }, "openai/whisper-large-v3": { id: "openai/whisper-large-v3", name: "Whisper 3 Large", description: "Open Whisper checkpoint for robust multilingual transcription and captioning", family: "whisper", attachment: false, reasoning: false, tool_call: false, release_date: "2024-10-01", last_updated: "2024-10-01", modalities: { input: ["audio"], output: ["text"] }, open_weights: true, limit: { context: 448, output: 4096 } }, "openai/gpt-5.5-pro": { id: "openai/gpt-5.5-pro", name: "GPT-5.5 Pro", description: "Highest-accuracy GPT-5.5 tier for slower, precision-heavy reasoning and coding", family: "gpt-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-12-01", release_date: "2026-04-23", last_updated: "2026-04-23", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "BrowseComp", score: 90.1, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 43.1, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 57.2, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 52.4, metric: "accuracy", dataset: "Tier 1-3", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 39.6, metric: "accuracy", dataset: "Tier 4", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GDPval", score: 82.3, metric: "wins or ties", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GeneBench", score: 33.2, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }] }, "openai/gpt-4-turbo": { id: "openai/gpt-4-turbo", name: "GPT-4 Turbo", description: "Compact GPT model for low-latency assistance and high-volume workloads", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: false, temperature: true, knowledge: "2023-12", release_date: "2023-11-06", last_updated: "2024-04-09", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 4096 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 21.5, metric: "index", source: "https://openrouter.ai/openai/gpt-4-turbo/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 31.9, metric: "percent correct", source: "https://openrouter.ai/openai/gpt-4-turbo/benchmarks", date: "2026-03-11" }] }, "openai/gpt-4": { id: "openai/gpt-4", name: "GPT-4", description: "GPT model for general reasoning, writing, coding, and tool-assisted tasks", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: false, temperature: true, knowledge: "2023-11", release_date: "2023-11-06", last_updated: "2024-04-09", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 8192, output: 8192 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 13.1, metric: "index", source: "https://openrouter.ai/openai/gpt-4/benchmarks", date: "2026-03-11" }] }, "openai/gpt-5-chat-latest": { id: "openai/gpt-5-chat-latest", name: "GPT-5 Chat (latest)", description: "Chat-tuned GPT model for conversational assistance, writing, and tool workflows", family: "gpt-codex", attachment: true, reasoning: true, tool_call: false, structured_output: true, temperature: true, knowledge: "2024-09-30", release_date: "2025-08-07", last_updated: "2025-08-07", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-5.6-sol": { id: "openai/gpt-5.6-sol", name: "GPT-5.6 Sol", description: "Frontier GPT-5.6 model for complex professional work, coding, and agentic workflows", family: "gpt-sol", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2026-02-16", release_date: "2026-07-09", last_updated: "2026-07-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 64.6, metric: "resolve rate", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Terminal-Bench", score: 88.8, metric: "success rate", version: "2.1", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "DeepSWE", score: 72.7, metric: "resolve rate", version: "1.1", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "GPQA Diamond", score: 94.6, metric: "accuracy", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "FrontierMath", score: 89, metric: "accuracy", dataset: "Tier 1-3", version: "v2", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "BrowseComp", score: 90.4, metric: "accuracy", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "OSWorld", score: 62.6, metric: "success rate", version: "2.0", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "MMMU Pro", score: 83, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Agents' Last Exam", score: 52.7, source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Toolathlon", score: 58, metric: "success rate", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Artificial Analysis Intelligence Index", score: 58.9, metric: "index score", variant: "max", version: "4.1", source: "https://artificialanalysis.ai/articles/gpt-5-6-has-landed", date: "2026-07-09" }, { name: "Artificial Analysis Coding Agent Index", score: 80, metric: "index score", harness: "Codex", variant: "max", version: "1.1", source: "https://artificialanalysis.ai/articles/gpt-5-6-has-landed", date: "2026-07-09" }] }, "openai/o3-pro": { id: "openai/o3-pro", name: "o3-pro", description: "High-effort o3 tier for difficult technical reasoning and careful answers", family: "o-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-05", release_date: "2025-06-10", last_updated: "2025-06-10", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 }, benchmarks: [{ name: "Aider Polyglot", score: 84.9, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-06-28" }] }, "openai/o3-deep-research": { id: "openai/o3-deep-research", name: "o3-deep-research", description: "Research model for long-horizon investigation, synthesis, and analytical reports", family: "o", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2024-05", release_date: "2024-06-26", last_updated: "2024-06-26", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 } }, "openai/gpt-5.4-pro": { id: "openai/gpt-5.4-pro", name: "GPT-5.4 Pro", description: "More exact GPT-5.4 tier for demanding professional reasoning and agent tasks", family: "gpt-pro", attachment: true, reasoning: true, tool_call: true, structured_output: false, temperature: false, knowledge: "2025-08-31", release_date: "2026-03-05", last_updated: "2026-03-05", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "GPQA Diamond", score: 94.4, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 42.7, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 58.7, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "BrowseComp", score: 89.3, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GDPval", score: 82, metric: "wins or ties", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 50, metric: "accuracy", dataset: "Tier 1-3", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 38, metric: "accuracy", dataset: "Tier 4", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "ARC-AGI-1", score: 94.5, metric: "accuracy", variant: "Verified", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "ARC-AGI-2", score: 83.3, metric: "accuracy", variant: "Verified", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FinanceAgent", score: 61.5, metric: "accuracy", version: "1.1", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GeneBench", score: 25.6, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }] }, "openai/gpt-5": { id: "openai/gpt-5", name: "GPT-5", description: "Original GPT-5 workhorse for reasoning, coding, writing, and tool workflows", family: "gpt", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-08-07", last_updated: "2025-08-07", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "Aider Polyglot", score: 88, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-08-23" }, { name: "SWE-Bench Pro", score: 41.78, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "openai/gpt-5.1-codex-mini": { id: "openai/gpt-5.1-codex-mini", name: "GPT-5.1 Codex mini", description: "Coding-optimized GPT model for repository edits, reviews, and agentic software work", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-11-13", last_updated: "2025-11-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-5-codex": { id: "openai/gpt-5-codex", name: "GPT-5-Codex", description: "Coding-optimized GPT model for repository edits, reviews, and agentic software work", family: "gpt-codex", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-09-15", last_updated: "2025-09-15", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 38.9, metric: "index", source: "https://openrouter.ai/openai/gpt-5-codex/benchmarks", date: "2026-06-01" }, { name: "SciCode", score: 40.9, metric: "percent correct", source: "https://openrouter.ai/openai/gpt-5-codex/benchmarks", date: "2026-06-01" }, { name: "Terminal-Bench Hard", score: 37.9, metric: "success rate", source: "https://openrouter.ai/openai/gpt-5-codex/benchmarks", date: "2026-06-01" }] }, "openai/gpt-5-mini": { id: "openai/gpt-5-mini", name: "GPT-5 Mini", description: "Small GPT-5 for responsive agents, coding help, and everyday automation", family: "gpt-mini", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-05-30", release_date: "2025-08-07", last_updated: "2025-08-07", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-5.6-luna": { id: "openai/gpt-5.6-luna", name: "GPT-5.6 Luna", description: "Cost-efficient GPT-5.6 model for fast, high-volume workloads", family: "gpt-luna", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2026-02-16", release_date: "2026-07-09", last_updated: "2026-07-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 62.7, metric: "resolve rate", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Terminal-Bench", score: 84.7, metric: "success rate", version: "2.1", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "DeepSWE", score: 67.2, metric: "resolve rate", version: "1.1", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "GPQA Diamond", score: 92.3, metric: "accuracy", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "FrontierMath", score: 78.6, metric: "accuracy", dataset: "Tier 1-3", version: "v2", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "BrowseComp", score: 83.3, metric: "accuracy", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "OSWorld", score: 45.6, metric: "success rate", version: "2.0", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "MMMU Pro", score: 78.4, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Agents' Last Exam", score: 50.3, source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Toolathlon", score: 53.4, metric: "success rate", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Artificial Analysis Intelligence Index", score: 51.2, metric: "index score", variant: "max", version: "4.1", source: "https://artificialanalysis.ai/articles/gpt-5-6-has-landed", date: "2026-07-09" }, { name: "Artificial Analysis Coding Agent Index", score: 74.6, metric: "index score", harness: "Codex", variant: "max", version: "1.1", source: "https://artificialanalysis.ai/articles/gpt-5-6-has-landed", date: "2026-07-09" }] }, "openai/gpt-5.3-codex": { id: "openai/gpt-5.3-codex", name: "GPT-5.3 Codex", description: "Coding-optimized GPT model for repository edits, reviews, and agentic software work", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2026-02-05", last_updated: "2026-02-05", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "SWE-Atlas Codebase QnA", score: 32.6, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 42.38, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 38.98, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }] }, "openai/gpt-oss-120b": { id: "openai/gpt-oss-120b", name: "GPT OSS 120B", description: "Open GPT reasoning model for self-hosted agents and controllable deployments", family: "gpt-oss", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2025-08-05", last_updated: "2025-08-05", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/openai/gpt-oss-120b" }] }, "openai/gpt-5.3-codex-spark": { id: "openai/gpt-5.3-codex-spark", name: "GPT-5.3 Codex Spark", description: "Coding-optimized GPT model for repository edits, reviews, and agentic software work", family: "gpt-codex-spark", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2026-02-05", last_updated: "2026-02-05", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 128e3, input: 1e5, output: 32e3 } }, "openai/gpt-4.1-nano": { id: "openai/gpt-4.1-nano", name: "GPT-4.1 nano", description: "Tiny GPT-4.1 option for classification, routing, and very high-volume tasks", family: "gpt-nano", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-04", release_date: "2025-04-14", last_updated: "2025-04-14", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 1047576, output: 32768 }, benchmarks: [{ name: "Aider Polyglot", score: 8.9, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-04-14" }] }, "openai/o1": { id: "openai/o1", name: "o1", description: "O-series reasoning model for hard analysis, math, coding, and planning", family: "o", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2023-09", release_date: "2024-12-05", last_updated: "2024-12-05", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 }, benchmarks: [{ name: "Aider Polyglot", score: 61.7, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2024-12-21" }] }, "openai/gpt-5.2": { id: "openai/gpt-5.2", name: "GPT-5.2", description: "Reliable GPT generation for broad coding, writing, and tool-assisted product work", family: "gpt", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2025-12-11", last_updated: "2025-12-11", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 29.94, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "openai/gpt-image-1": { id: "openai/gpt-image-1", name: "GPT-Image-1", description: "OpenAI image model for production generation, edits, and brand-safe visual workflows", family: "gpt-image", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2025-04-24", last_updated: "2025-04-24", modalities: { input: ["text", "image"], output: ["image"] }, open_weights: false, limit: { context: 0, output: 0 } }, "openai/gpt-realtime-2.1": { id: "openai/gpt-realtime-2.1", name: "GPT-Realtime-2.1", description: "Realtime speech-to-speech model with configurable reasoning, tool use, and robust voice-agent behavior", family: "gpt", attachment: true, reasoning: true, tool_call: true, structured_output: false, temperature: false, knowledge: "2024-09-30", release_date: "2026-07-06", last_updated: "2026-07-06", modalities: { input: ["text", "audio", "image"], output: ["text", "audio"] }, open_weights: false, limit: { context: 128e3, input: 96e3, output: 32e3 } }, "openai/gpt-5.4-mini": { id: "openai/gpt-5.4-mini", name: "GPT-5.4 mini", description: "Strong small GPT for coding subagents, quick tool use, and high-volume work", family: "gpt-mini", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2026-03-17", last_updated: "2026-03-17", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 54.4, metric: "resolve rate", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Terminal-Bench", score: 60, metric: "accuracy", variant: "reasoning effort xhigh", version: "2.0", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "MCP Atlas", score: 57.7, metric: "score", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Toolathlon", score: 42.9, metric: "score", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "\u03C4\xB2-Bench Telecom", score: 93.4, metric: "accuracy", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "GPQA Diamond", score: 88, metric: "accuracy", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Humanity's Last Exam", score: 41.5, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Humanity's Last Exam", score: 28.2, metric: "accuracy", variant: "without tools", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OSWorld-Verified", score: 72.1, metric: "success rate", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "MMMU Pro", score: 78, metric: "accuracy", variant: "with Python", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "MMMU Pro", score: 76.6, metric: "accuracy", variant: "without tools", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OmniDocBench", score: 0.1263, metric: "overall edit distance", variant: "reasoning effort none", version: "1.5", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OpenAI MRCR", score: 47.7, metric: "accuracy", variant: "8-needle, 64K-128K", version: "v2", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OpenAI MRCR", score: 33.6, metric: "accuracy", variant: "8-needle, 128K-256K", version: "v2", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Graphwalks", score: 76.3, metric: "accuracy", variant: "BFS, 0-128K", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Graphwalks", score: 71.5, metric: "accuracy", variant: "parents, 0-128K", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }] }, "openai/gpt-4o-2024-05-13": { id: "openai/gpt-4o-2024-05-13", name: "GPT-4o (2024-05-13)", description: "GPT model for general reasoning, writing, coding, and tool-assisted tasks", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2023-09", release_date: "2024-05-13", last_updated: "2024-05-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 4096 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 24.2, metric: "index", source: "https://openrouter.ai/openai/gpt-4o-2024-05-13/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 30.9, metric: "percent correct", source: "https://openrouter.ai/openai/gpt-4o-2024-05-13/benchmarks", date: "2026-03-11" }] }, "openai/o3": { id: "openai/o3", name: "o3", description: "Deliberate o-series reasoner for hard math, coding, and multi-step analysis", family: "o", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-05", release_date: "2025-04-16", last_updated: "2025-04-16", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 }, benchmarks: [{ name: "Aider Polyglot", score: 81.3, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-06-25" }] }, "openai/o4-mini-deep-research": { id: "openai/o4-mini-deep-research", name: "o4-mini-deep-research", description: "Research model for long-horizon investigation, synthesis, and analytical reports", family: "o-mini", attachment: true, reasoning: true, tool_call: true, temperature: false, knowledge: "2024-05", release_date: "2024-06-26", last_updated: "2024-06-26", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 } }, "openai/gpt-5-pro": { id: "openai/gpt-5-pro", name: "GPT-5 Pro", description: "Higher-accuracy GPT-5 tier for tough analysis, coding reviews, and planning", family: "gpt-pro", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-10-06", last_updated: "2025-10-06", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 272e3 } }, "openai/gpt-5.5": { id: "openai/gpt-5.5", name: "GPT-5.5", description: "Default frontier GPT for coding, computer use, research, and knowledge work", family: "gpt", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-12-01", release_date: "2026-04-23", last_updated: "2026-04-23", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 58.6, metric: "resolve rate", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "Terminal-Bench", score: 78.2, metric: "success rate", harness: "Terminus-2", version: "2.1", source: "https://www.anthropic.com/news/claude-opus-4-8", date: "2026-05-28" }, { name: "SWE-Atlas Codebase QnA", score: 45.43, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 44.79, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 42.59, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "Artificial Analysis Coding Agent Index", score: 65.3, metric: "average pass@1", harness: "Codex", variant: "xhigh", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 80.8, metric: "pass@1", harness: "Codex", variant: "xhigh", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 30.9, metric: "pass@1", harness: "Codex", variant: "xhigh", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 84.1, metric: "pass@1", harness: "Codex", variant: "xhigh", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Artificial Analysis Coding Agent Index", score: 60.4, metric: "average pass@1", harness: "Codex", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 79.1, metric: "pass@1", harness: "Codex", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 26.2, metric: "pass@1", harness: "Codex", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 75.8, metric: "pass@1", harness: "Codex", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Artificial Analysis Coding Agent Index", score: 57.8, metric: "average pass@1", harness: "Cursor CLI", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 75, metric: "pass@1", harness: "Cursor CLI", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 24.9, metric: "pass@1", harness: "Cursor CLI", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 73.4, metric: "pass@1", harness: "Cursor CLI", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 82.7, metric: "success rate", version: "2.0", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GPQA Diamond", score: 93.6, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 41.4, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 52.2, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "OSWorld-Verified", score: 78.7, metric: "success rate", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "BrowseComp", score: 84.4, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "MMMU Pro", score: 81.2, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "ARC-AGI-2", score: 85, metric: "accuracy", variant: "Verified", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 51.7, metric: "accuracy", dataset: "Tier 1-3", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 35.4, metric: "accuracy", dataset: "Tier 4", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GDPval", score: 84.9, metric: "wins or ties", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "MCP Atlas", score: 75.3, metric: "success rate", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Toolathlon", score: 55.6, metric: "success rate", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "\u03C4\xB2-Bench Telecom", score: 98, metric: "success rate", variant: "original prompts", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }] }, "openai/gpt-5.5-instant": { id: "openai/gpt-5.5-instant", name: "GPT-5.5 Instant", description: "Compact GPT model for low-latency assistance and high-volume workloads", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-12-01", release_date: "2026-05-05", last_updated: "2026-05-28", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 4e5, output: 128e3 } }, "openai/gpt-5.2-codex": { id: "openai/gpt-5.2-codex", name: "GPT-5.2 Codex", description: "Code-specialist GPT for repository edits, reviews, and long-running software agents", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2025-12-11", last_updated: "2025-12-11", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 41.04, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "openai/gpt-4.1": { id: "openai/gpt-4.1", name: "GPT-4.1", description: "Long-lived GPT workhorse for coding, instruction following, and production apps", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-04", release_date: "2025-04-14", last_updated: "2025-04-14", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1047576, output: 32768 }, benchmarks: [{ name: "Aider Polyglot", score: 52.4, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-04-14" }] }, "openai/gpt-4o-2024-08-06": { id: "openai/gpt-4o-2024-08-06", name: "GPT-4o (2024-08-06)", description: "GPT model for general reasoning, writing, coding, and tool-assisted tasks", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2023-09", release_date: "2024-08-06", last_updated: "2024-08-06", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 }, benchmarks: [{ name: "Aider Polyglot", score: 23.1, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2024-12-30" }, { name: "Artificial Analysis Coding Index", score: 16.6, metric: "index", source: "https://openrouter.ai/openai/gpt-4o-2024-08-06/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 33.1, metric: "percent correct", source: "https://openrouter.ai/openai/gpt-4o-2024-08-06/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 8.3, metric: "success rate", source: "https://openrouter.ai/openai/gpt-4o-2024-08-06/benchmarks", date: "2026-03-11" }] }, "openai/gpt-5.6-terra": { id: "openai/gpt-5.6-terra", name: "GPT-5.6 Terra", description: "Balanced GPT-5.6 model for capable, cost-efficient everyday work", family: "gpt-terra", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2026-02-16", release_date: "2026-07-09", last_updated: "2026-07-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 63.4, metric: "resolve rate", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Terminal-Bench", score: 87.4, metric: "success rate", version: "2.1", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "DeepSWE", score: 69.6, metric: "resolve rate", version: "1.1", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "GPQA Diamond", score: 92.9, metric: "accuracy", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "FrontierMath", score: 84.9, metric: "accuracy", dataset: "Tier 1-3", version: "v2", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "BrowseComp", score: 87.5, metric: "accuracy", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "OSWorld", score: 50.2, metric: "success rate", version: "2.0", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "MMMU Pro", score: 80.7, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Agents' Last Exam", score: 50.4, source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Toolathlon", score: 53.1, metric: "success rate", source: "https://openai.com/index/gpt-5-6/", date: "2026-07-09" }, { name: "Artificial Analysis Intelligence Index", score: 55, metric: "index score", variant: "max", version: "4.1", source: "https://artificialanalysis.ai/articles/gpt-5-6-has-landed", date: "2026-07-09" }, { name: "Artificial Analysis Coding Agent Index", score: 77.4, metric: "index score", harness: "Codex", variant: "max", version: "1.1", source: "https://artificialanalysis.ai/articles/gpt-5-6-has-landed", date: "2026-07-09" }] }, "openai/o4-mini": { id: "openai/o4-mini", name: "o4-mini", description: "Fast o-series model for compact reasoning, coding, and tool use", family: "o-mini", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-05", release_date: "2025-04-16", last_updated: "2025-04-16", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 }, benchmarks: [{ name: "Aider Polyglot", score: 72, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-04-16" }] }, "openai/gpt-5.4": { id: "openai/gpt-5.4", name: "GPT-5.4", description: "Agent-ready GPT for coding and computer-use workflows at a lower cost", family: "gpt", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2026-03-05", last_updated: "2026-03-05", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 105e4, input: 922e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 59.1, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }, { name: "SWE-Atlas Codebase QnA", score: 40.8, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Codebase QnA", score: 36.3, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 44.29, metric: "score", harness: "Codex", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 44.36, metric: "score", harness: "Codex CLI", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "SWE-Atlas Test Writing", score: 40, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }, { name: "Artificial Analysis Coding Agent Index", score: 53.6, metric: "average pass@1", harness: "Codex", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 72.4, metric: "pass@1", harness: "Codex", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 18.4, metric: "pass@1", harness: "Codex", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 69.8, metric: "pass@1", harness: "Codex", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Artificial Analysis Coding Agent Index", score: 52.2, metric: "average pass@1", harness: "Cursor CLI", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 72.9, metric: "pass@1", harness: "Cursor CLI", variant: "medium", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 18.9, metric: "pass@1", harness: "Cursor CLI", variant: "medium", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 64.7, metric: "pass@1", harness: "Cursor CLI", variant: "medium", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 75.1, metric: "success rate", version: "2.0", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GPQA Diamond", score: 92.8, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 39.8, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "Humanity's Last Exam", score: 52.1, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "OSWorld-Verified", score: 75, metric: "success rate", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "BrowseComp", score: 82.7, metric: "accuracy", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "GDPval", score: 83, metric: "wins or ties", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "ARC-AGI-2", score: 73.3, metric: "accuracy", variant: "Verified", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 47.6, metric: "accuracy", dataset: "Tier 1-3", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "FrontierMath", score: 27.1, metric: "accuracy", dataset: "Tier 4", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }, { name: "MMMU Pro", score: 81.2, metric: "accuracy", variant: "no tools", source: "https://openai.com/index/introducing-gpt-5-5/", date: "2026-04-23" }] }, "openai/o3-mini": { id: "openai/o3-mini", name: "o3-mini", description: "Smaller o-series reasoner for economical coding, math, and planning tasks", family: "o-mini", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-05", release_date: "2024-12-20", last_updated: "2025-01-29", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 1e5 }, benchmarks: [{ name: "Aider Polyglot", score: 60.4, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-01-31" }] }, "openai/gpt-4o-2024-11-20": { id: "openai/gpt-4o-2024-11-20", name: "GPT-4o (2024-11-20)", description: "GPT model for general reasoning, writing, coding, and tool-assisted tasks", family: "gpt", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2023-09", release_date: "2024-11-20", last_updated: "2024-11-20", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 }, benchmarks: [{ name: "Aider Polyglot", score: 18.2, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2024-12-30" }, { name: "Artificial Analysis Coding Index", score: 16.7, metric: "index", source: "https://openrouter.ai/openai/gpt-4o-2024-11-20/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 33.3, metric: "percent correct", source: "https://openrouter.ai/openai/gpt-4o-2024-11-20/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 8.3, metric: "success rate", source: "https://openrouter.ai/openai/gpt-4o-2024-11-20/benchmarks", date: "2026-03-11" }] }, "openai/gpt-oss-safeguard-120b": { id: "openai/gpt-oss-safeguard-120b", name: "GPT OSS Safeguard 120B", description: "Safety model for policy screening, moderation, and risk-aware routing workflows", family: "gpt-oss", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2025-10-29", last_updated: "2025-10-29", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/openai/gpt-oss-safeguard-120b" }] }, "openai/gpt-realtime-whisper": { id: "openai/gpt-realtime-whisper", name: "GPT Realtime Whisper", description: "Streaming speech-to-text model for low-latency transcript deltas from live audio", family: "whisper", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2026-05-07", last_updated: "2026-05-07", modalities: { input: ["audio"], output: ["text"] }, open_weights: false, limit: { context: 0, output: 0 } }, "openai/gpt-5.2-chat-latest": { id: "openai/gpt-5.2-chat-latest", name: "GPT-5.2 Chat", description: "Chat-tuned GPT model for conversational assistance, writing, and tool workflows", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2025-12-11", last_updated: "2025-12-11", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 } }, "openai/gpt-5.2-pro": { id: "openai/gpt-5.2-pro", name: "GPT-5.2 Pro", description: "Higher-accuracy GPT-5.2 variant for tougher reasoning and review workflows", family: "gpt-pro", attachment: true, reasoning: true, tool_call: true, structured_output: false, temperature: false, knowledge: "2025-08-31", release_date: "2025-12-11", last_updated: "2025-12-11", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-5.1-chat-latest": { id: "openai/gpt-5.1-chat-latest", name: "GPT-5.1 Chat", description: "Chat-tuned GPT-5.1 for polished assistants, writing, and product conversations", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-11-13", last_updated: "2025-11-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 } }, "openai/gpt-image-2": { id: "openai/gpt-image-2", name: "GPT-Image-2", description: "Image model for prompt-driven generation, editing, and visual design workflows", family: "gpt-image", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2026-04-21", last_updated: "2026-04-21", modalities: { input: ["text", "image"], output: ["image"] }, open_weights: false, limit: { context: 0, output: 0 } }, "openai/gpt-5.1-codex-max": { id: "openai/gpt-5.1-codex-max", name: "GPT-5.1 Codex Max", description: "Coding-optimized GPT model for repository edits, reviews, and agentic software work", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-11-13", last_updated: "2025-11-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-5.1-codex": { id: "openai/gpt-5.1-codex", name: "GPT-5.1 Codex", description: "Codex GPT for repository edits, code review, and practical software agents", family: "gpt-codex", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-11-13", last_updated: "2025-11-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "openai/gpt-4.1-mini": { id: "openai/gpt-4.1-mini", name: "GPT-4.1 mini", description: "Affordable GPT-4.1 lane for fast coding help and structured extraction", family: "gpt-mini", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-04", release_date: "2025-04-14", last_updated: "2025-04-14", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1047576, output: 32768 }, benchmarks: [{ name: "Aider Polyglot", score: 32.4, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-04-14" }] }, "openai/gpt-oss-20b": { id: "openai/gpt-oss-20b", name: "GPT OSS 20B", description: "Open GPT reasoning model for self-hosted agents and controllable deployments", family: "gpt-oss", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2025-08-05", last_updated: "2025-08-05", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/openai/gpt-oss-20b" }] }, "openai/whisper-large-v3-turbo": { id: "openai/whisper-large-v3-turbo", name: "Whisper Large v3 Turbo", description: "Speech transcription model for accurate audio-to-text and captioning workflows", family: "whisper", attachment: false, reasoning: false, tool_call: false, release_date: "2024-10-01", last_updated: "2024-10-01", modalities: { input: ["audio"], output: ["text"] }, open_weights: true, limit: { context: 448, output: 448 } }, "openai/gpt-5.4-nano": { id: "openai/gpt-5.4-nano", name: "GPT-5.4 nano", description: "Cheapest GPT-5.4 lane for simple routing, extraction, and bulk automation", family: "gpt-nano", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2025-08-31", release_date: "2026-03-17", last_updated: "2026-03-17", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Pro", score: 52.4, metric: "resolve rate", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Terminal-Bench", score: 46.3, metric: "accuracy", variant: "reasoning effort xhigh", version: "2.0", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "MCP Atlas", score: 56.1, metric: "score", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Toolathlon", score: 35.5, metric: "score", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "\u03C4\xB2-Bench Telecom", score: 92.5, metric: "accuracy", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "GPQA Diamond", score: 82.8, metric: "accuracy", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Humanity's Last Exam", score: 37.7, metric: "accuracy", variant: "with tools", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Humanity's Last Exam", score: 24.3, metric: "accuracy", variant: "without tools", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OSWorld-Verified", score: 39, metric: "success rate", variant: "reasoning effort xhigh", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "MMMU Pro", score: 69.5, metric: "accuracy", variant: "with Python", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "MMMU Pro", score: 66.1, metric: "accuracy", variant: "without tools", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OmniDocBench", score: 0.2419, metric: "overall edit distance", variant: "reasoning effort none", version: "1.5", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OpenAI MRCR", score: 44.2, metric: "accuracy", variant: "8-needle, 64K-128K", version: "v2", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "OpenAI MRCR", score: 33.1, metric: "accuracy", variant: "8-needle, 128K-256K", version: "v2", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Graphwalks", score: 73.4, metric: "accuracy", variant: "BFS, 0-128K", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }, { name: "Graphwalks", score: 50.8, metric: "accuracy", variant: "parents, 0-128K", source: "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/", date: "2026-03-17" }] }, "openai/gpt-5.1": { id: "openai/gpt-5.1", name: "GPT-5.1", description: "Sharper GPT-5 generation for coding, product work, and tool-assisted tasks", family: "gpt", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, knowledge: "2024-09-30", release_date: "2025-11-13", last_updated: "2025-11-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 4e5, input: 272e3, output: 128e3 } }, "xai/grok-4.6": { id: "xai/grok-4.6", name: "Grok 4.6", description: "xAI's frontier model for long-running agents, coding, knowledge work, and visual projects", family: "grok", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-02-01", release_date: "2026-08-12", last_updated: "2026-08-12", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 5e5, output: 5e5 } }, "xai/grok-4.5": { id: "xai/grok-4.5", name: "Grok 4.5", description: "xAI's Grok model for chat, coding, agentic tools, and lower hallucination risk", family: "grok", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-07-08", last_updated: "2026-07-08", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 5e5, output: 5e5 }, benchmarks: [{ name: "SWE-Bench Pro", score: 64.7, metric: "resolve rate", source: "https://x.ai/news/grok-4-5", date: "2026-07-08" }, { name: "SWE-Bench Multilingual", score: 78, metric: "resolve rate", source: "https://x.ai/news/grok-4-5", date: "2026-07-08" }, { name: "Terminal-Bench", score: 83.3, metric: "success rate", version: "2.1", source: "https://x.ai/news/grok-4-5", date: "2026-07-08" }, { name: "DeepSWE", score: 62, metric: "resolve rate", version: "1.0", source: "https://x.ai/news/grok-4-5", date: "2026-07-08" }, { name: "DeepSWE", score: 53, metric: "resolve rate", harness: "mini-swe-agent", version: "1.1", source: "https://x.ai/news/grok-4-5", date: "2026-07-08" }, { name: "SWE Marathon", score: 29, metric: "pass@1", source: "https://x.ai/news/grok-4-5", date: "2026-07-08" }] }, "xai/grok-4.20-0309-non-reasoning": { id: "xai/grok-4.20-0309-non-reasoning", name: "Grok 4.20 (Non-Reasoning)", description: "Grok model for agentic tool use, reasoning, coding, and live assistance", family: "grok", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, release_date: "2026-03-09", last_updated: "2026-03-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 3e4 } }, "xai/grok-imagine-video-1.5": { id: "xai/grok-imagine-video-1.5", name: "Grok Imagine Video 1.5", description: "Video model for image-to-video generation, editing, and extension workflows", family: "grok", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2026-05-30", last_updated: "2026-05-30", modalities: { input: ["text", "image", "video"], output: ["video"] }, open_weights: false, limit: { context: 1024, output: 0 } }, "xai/grok-4.1-fast": { id: "xai/grok-4.1-fast", name: "Grok 4.1 Fast", description: "xAI's fast agentic tool-calling model with a 2M context window; non-reasoning variant for low-latency responses", family: "grok", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, release_date: "2025-11-19", last_updated: "2025-11-19", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e6, output: 3e4 } }, "xai/grok-4.20-0309-reasoning": { id: "xai/grok-4.20-0309-reasoning", name: "Grok 4.20 (Reasoning)", description: "Reasoning Grok for document-heavy analysis and long-horizon tool use", family: "grok", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-03-09", last_updated: "2026-03-09", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 3e4 } }, "xai/grok-imagine-image-2.0": { id: "xai/grok-imagine-image-2.0", name: "Grok Imagine Image 2.0", description: "Image model for prompt-driven generation, editing, and visual design workflows", family: "grok", attachment: true, reasoning: false, tool_call: false, temperature: false, release_date: "2026-08-07", last_updated: "2026-08-07", modalities: { input: ["text", "image"], output: ["image"] }, open_weights: false, limit: { context: 8e3, output: 0 } }, "xai/grok-build-0.1": { id: "xai/grok-build-0.1", name: "Grok Build 0.1", description: "Fast Grok coding model tuned for agentic engineering and iterative edits", family: "grok-build", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-16", last_updated: "2026-04-16", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 256e3, output: 256e3 } }, "xai/grok-4.3": { id: "xai/grok-4.3", name: "Grok 4.3", description: "xAI's default Grok for chat, coding, agentic tools, and lower hallucination risk", family: "grok", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-17", last_updated: "2026-04-17", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 3e4 }, benchmarks: [{ name: "Artificial Analysis Intelligence Index", score: 53, metric: "index score", version: "4.0", source: "https://artificialanalysis.ai/articles/xai-launches-grok-4-3-with-improved-agentic-performance-and-lower-pricing", date: "2026-04-30" }, { name: "GDPval-AA", score: 1500, metric: "Elo", source: "https://artificialanalysis.ai/articles/xai-launches-grok-4-3-with-improved-agentic-performance-and-lower-pricing", date: "2026-04-30" }, { name: "\u03C4\xB2-Bench Telecom", score: 98, metric: "success rate", source: "https://artificialanalysis.ai/articles/xai-launches-grok-4-3-with-improved-agentic-performance-and-lower-pricing", date: "2026-04-30" }, { name: "IFBench", score: 81, metric: "accuracy", source: "https://artificialanalysis.ai/articles/xai-launches-grok-4-3-with-improved-agentic-performance-and-lower-pricing", date: "2026-04-30" }] }, "meituan/longcat-2.0": { id: "meituan/longcat-2.0", name: "LongCat-2.0", description: "Meituan LongCat-2.0, a reasoning model with tool calling and a 1M-token context window", family: "longcat", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-30", last_updated: "2026-06-30", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 131072 }, benchmarks: [{ name: "SWE-Bench Pro", score: 59.5, metric: "resolve rate", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }, { name: "SWE-Bench Multilingual", score: 77.3, metric: "resolve rate", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }, { name: "Terminal-Bench", score: 70.8, metric: "success rate", version: "2.1", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }, { name: "GPQA Diamond", score: 88.9, metric: "accuracy", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }, { name: "BrowseComp", score: 79.9, metric: "accuracy", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }, { name: "IFEval", score: 90, metric: "accuracy", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }, { name: "FORTE", score: 73.2, metric: "success rate", source: "https://github.com/meituan-longcat/longcat-2.0", date: "2026-06-30" }] }, "swiss-ai/apertus-70b": { id: "swiss-ai/apertus-70b", name: "Apertus 70B", description: "Fully open 70B multilingual LLM supporting 1800+ languages with 65K context. Trained on 15T tokens of compliant open data. Apache 2.0, EU AI Act compliant.", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-09", release_date: "2025-09-02", last_updated: "2025-09-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 65536, output: 8192 }, license: "Apache-2.0", links: [{ label: "Paper", url: "https://arxiv.org/abs/2509.14233", type: "paper" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509" }] }, "swiss-ai/apertus-8b": { id: "swiss-ai/apertus-8b", name: "Apertus 8B", description: "Fully open 8B multilingual LLM supporting 1800+ languages with 65K context. Trained on compliant open data. Apache 2.0, EU AI Act compliant.", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-09", release_date: "2025-09-02", last_updated: "2025-09-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 65536, output: 8192 }, license: "Apache-2.0", links: [{ label: "Paper", url: "https://arxiv.org/abs/2509.14233", type: "paper" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509" }] }, "stepfun/step-3.5-flash-2603": { id: "stepfun/step-3.5-flash-2603", name: "Step 3.5 Flash 2603", description: "StepFun flash model for efficient multimodal reasoning, coding, and tool use", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-01", release_date: "2026-04-02", last_updated: "2026-04-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, input: 256e3, output: 256e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/stepfun-ai/Step-3.5-Flash" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 34.6, metric: "index", source: "https://openrouter.ai/stepfun/step-3.5-flash/benchmarks", date: "2026-06-02" }, { name: "SciCode", score: 38.5, metric: "percent correct", source: "https://openrouter.ai/stepfun/step-3.5-flash/benchmarks", date: "2026-06-02" }, { name: "Terminal-Bench Hard", score: 32.6, metric: "success rate", source: "https://openrouter.ai/stepfun/step-3.5-flash/benchmarks", date: "2026-06-02" }] }, "stepfun/step-3.5-flash": { id: "stepfun/step-3.5-flash", name: "Step 3.5 Flash", description: "StepFun flash lane for quick multimodal reasoning and coding assistance", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-01", release_date: "2026-01-29", last_updated: "2026-02-13", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, input: 256e3, output: 256e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/stepfun-ai/Step-3.5-Flash" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 31.6, metric: "index", source: "https://openrouter.ai/stepfun/step-3.5-flash/benchmarks", date: "2026-06-02" }, { name: "SciCode", score: 40.4, metric: "percent correct", source: "https://openrouter.ai/stepfun/step-3.5-flash/benchmarks", date: "2026-06-02" }, { name: "Terminal-Bench Hard", score: 27.3, metric: "success rate", source: "https://openrouter.ai/stepfun/step-3.5-flash/benchmarks", date: "2026-06-02" }, { name: "SWE-Bench Verified", score: 74.4, metric: "resolved", source: "https://arxiv.org/abs/2602.10604" }] }, "stepfun/step-3.7-flash": { id: "stepfun/step-3.7-flash", name: "Step 3.7 Flash", description: "Newer StepFun flash model for faster agents, coding, and multimodal prompts", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2026-03-01", release_date: "2026-05-29", last_updated: "2026-05-29", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 256e3, input: 256e3, output: 256e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/stepfun-ai/Step-3.7-Flash" }], benchmarks: [{ name: "SWE-Bench Pro", score: 56.3, metric: "resolve rate", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "SWE-Bench Verified", score: 76.5, metric: "resolved", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "Terminal-Bench", score: 59.6, metric: "success rate", version: "2.1", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "Humanity's Last Exam", score: 47.2, metric: "accuracy", variant: "with tools", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "BrowseComp", score: 75.8, metric: "accuracy", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "Toolathlon", score: 49.5, metric: "success rate", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "GDPval", score: 45.8, metric: "wins or ties", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "ClawEval", score: 67.1, metric: "pass^3", version: "1.1", source: "https://static.stepfun.com/blog/step-3.7-flash/", date: "2026-05-29" }, { name: "Artificial Analysis Coding Index", score: 37.1, metric: "index", source: "https://openrouter.ai/stepfun/step-3.7-flash/benchmarks", date: "2026-06-15" }, { name: "SciCode", score: 40, metric: "percent correct", source: "https://openrouter.ai/stepfun/step-3.7-flash/benchmarks", date: "2026-06-15" }, { name: "Terminal-Bench Hard", score: 35.6, metric: "success rate", source: "https://openrouter.ai/stepfun/step-3.7-flash/benchmarks", date: "2026-06-15" }] }, "sarvam/sarvam-105b": { id: "sarvam/sarvam-105b", name: "Sarvam 105B", description: "Flagship Indian-language reasoning model for enterprise multilingual applications", family: "sarvam", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-09-01", last_updated: "2025-09-01", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 } }, "sarvam/sarvam-30b": { id: "sarvam/sarvam-30b", name: "Sarvam 30B", description: "Efficient Indian-language reasoning model for chat, coding, and multilingual work", family: "sarvam", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-18", last_updated: "2026-02-18", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 } }, "trendyol/asure-12b": { id: "trendyol/asure-12b", name: "Trendyol Asure 12B", description: "Turkish-language multimodal instruct model built on Gemma 3 12B for e-commerce text, chat, and image-text tasks", family: "gemma", attachment: true, reasoning: false, tool_call: false, temperature: true, release_date: "2026-02-19", last_updated: "2026-02-20", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 131072 }, license: "Gemma", weights: [{ label: "Hugging Face", url: "https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B" }] }, "deepreinforce/ornith-1.0-35b": { id: "deepreinforce/ornith-1.0-35b", name: "Ornith 1.0 35B", description: "Large coding-reasoning model for agentic software tasks and RL search", family: "ornith", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-25", last_updated: "2026-06-25", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144 }, license: "MIT", links: [{ label: "Model card", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B", type: "model_card" }, { label: "Announcement", url: "https://deep-reinforce.com/ornith_1_0.html", type: "announcement" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 75.6, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }, { name: "SWE-Bench Pro", score: 50.4, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }, { name: "SWE-Bench Multilingual", score: 69.3, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }, { name: "Terminal-Bench 2.1", score: 64.2, metric: "percent", variant: "Terminus-2", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }, { name: "Terminal-Bench 2.1", score: 62.8, metric: "percent", variant: "Claude Code", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }, { name: "NL2Repo", score: 34.6, metric: "percent", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }, { name: "Claw-eval", score: 69.8, metric: "percent", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B" }] }, "deepreinforce/ornith-1.0-397b": { id: "deepreinforce/ornith-1.0-397b", name: "Ornith 1.0 397B", description: "Large coding-reasoning model for agentic software tasks and RL search", family: "ornith", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-25", last_updated: "2026-06-25", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144 }, license: "MIT", links: [{ label: "Model card", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B", type: "model_card" }, { label: "Announcement", url: "https://deep-reinforce.com/ornith_1_0.html", type: "announcement" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { label: "Hugging Face (FP8)", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B-FP8", quantization: "fp8" }], benchmarks: [{ name: "SWE-Bench Verified", score: 82.4, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { name: "SWE-Bench Pro", score: 62.2, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { name: "SWE-Bench Multilingual", score: 78.9, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { name: "Terminal-Bench 2.1", score: 77.5, metric: "percent", variant: "Terminus-2", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { name: "Terminal-Bench 2.1", score: 78.2, metric: "percent", variant: "Claude Code", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { name: "NL2Repo", score: 48.2, metric: "percent", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }, { name: "Claw-eval", score: 77.1, metric: "percent", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-397B" }] }, "deepreinforce/ornith-1.0-31b": { id: "deepreinforce/ornith-1.0-31b", name: "Ornith 1.0 31B", description: "Open coding-reasoning model for repository tasks and self-improving agents", family: "ornith", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-25", last_updated: "2026-06-25", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144 }, license: "MIT", links: [{ label: "Announcement", url: "https://deep-reinforce.com/ornith_1_0.html", type: "announcement" }] }, "deepreinforce/ornith-1.0-9b": { id: "deepreinforce/ornith-1.0-9b", name: "Ornith 1.0 9B", description: "Open coding-reasoning model for repository tasks and self-improving agents", family: "ornith", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-25", last_updated: "2026-06-25", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144 }, license: "MIT", links: [{ label: "Model card", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B", type: "model_card" }, { label: "Announcement", url: "https://deep-reinforce.com/ornith_1_0.html", type: "announcement" }], weights: [{ label: "Hugging Face", url: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 69.4, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }, { name: "SWE-Bench Pro", score: 42.9, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }, { name: "SWE-Bench Multilingual", score: 52, metric: "percent resolved", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }, { name: "Terminal-Bench 2.1", score: 43.1, metric: "percent", variant: "Terminus-2", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }, { name: "Terminal-Bench 2.1", score: 40.6, metric: "percent", variant: "Claude Code", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }, { name: "NL2Repo", score: 27.2, metric: "percent", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }, { name: "Claw-eval", score: 63.1, metric: "percent", source: "https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B" }] }, "xiaomi/mimo-v2.5-pro-ultraspeed": { id: "xiaomi/mimo-v2.5-pro-ultraspeed", name: "MiMo-V2.5-Pro-UltraSpeed", description: "MiMo pro model for strong multimodal reasoning and agent execution", family: "mimo", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-12", release_date: "2026-06-08", last_updated: "2026-06-09", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash" }] }, "xiaomi/mimo-v2.5": { id: "xiaomi/mimo-v2.5", name: "MiMo-V2.5", description: "Open MiMo model for multimodal coding agents and long-context automation", family: "mimo", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-12", release_date: "2026-04-22", last_updated: "2026-04-22", modalities: { input: ["text", "image", "audio", "video"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/XiaomiMiMo/MiMo-V2.5" }] }, "xiaomi/mimo-v2-pro": { id: "xiaomi/mimo-v2-pro", name: "MiMo-V2-Pro", description: "Earlier MiMo Pro model for multimodal agents, reasoning, and code tasks", family: "mimo", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-12", release_date: "2026-03-18", last_updated: "2026-03-18", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 131072 } }, "xiaomi/mimo-v2-omni": { id: "xiaomi/mimo-v2-omni", name: "MiMo-V2-Omni", description: "MiMo omni model for text, image, video, audio, and agents", family: "mimo", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-12", release_date: "2026-03-18", last_updated: "2026-03-18", modalities: { input: ["text", "image", "audio", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 262144, output: 131072 } }, "xiaomi/mimo-v2-flash": { id: "xiaomi/mimo-v2-flash", name: "MiMo-V2-Flash", description: "MiMo flash model for fast multimodal assistance and agent workflows", family: "mimo", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-12-01", release_date: "2025-12-16", last_updated: "2026-02-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash" }] }, "xiaomi/mimo-v2.5-pro": { id: "xiaomi/mimo-v2.5-pro", name: "MiMo-V2.5-Pro", description: "Stronger MiMo Pro tier for multimodal reasoning and coding-agent execution", family: "mimo", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-12", release_date: "2026-04-22", last_updated: "2026-04-22", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro" }], benchmarks: [{ name: "SWE-Bench Verified", score: 78.9, metric: "resolved", source: "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro" }, { name: "SWE-Bench Pro", score: 57.2, metric: "resolve rate", source: "https://mimo.xiaomi.com/mimo-v2-5-pro/", date: "2026-04-22" }, { name: "GPQA Diamond", score: 86.6, metric: "accuracy", source: "https://mimo.xiaomi.com/mimo-v2-5-pro/", date: "2026-04-22" }] }, "alibaba/qwen3-vl-plus": { id: "alibaba/qwen3-vl-plus", name: "Qwen3-VL Plus", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-09-23", last_updated: "2025-09-23", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 262144, output: 32768 } }, "alibaba/qwen3.8-max-preview": { id: "alibaba/qwen3.8-max-preview", name: "Qwen3.8 Max Preview", description: "Preview Qwen flagship for million-token multimodal reasoning and long-horizon agentic workflows", family: "qwen", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-07-19", last_updated: "2026-07-19", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 131072 }, benchmarks: [{ name: "Terminal-Bench", score: 86.6, metric: "accuracy", variant: "xhigh", version: "2.1", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "SWE-Bench Pro", score: 67.7, metric: "resolve rate", harness: "Claude Code", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "DeepSWE", score: 56.6, metric: "resolve rate", harness: "Claude Code", variant: "xhigh", version: "1.1", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "NL2Repo", score: 55.9, metric: "resolve rate", harness: "Claude Code", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "FrontierSWE", score: 73.5, metric: "dominance score", harness: "Claude Code", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "MLS-Bench-Lite", score: 41, metric: "score", harness: "Claude Code", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "AutomationBench", score: 27.3, metric: "pass@1", variant: "xhigh", dataset: "600-task public subset", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "Toolathlon Verified", score: 72.5, metric: "pass@1", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "WideSearch", score: 81.9, metric: "F1", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "Humanity's Last Exam", score: 56.2, metric: "accuracy", variant: "xhigh, with tools", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "GPQA Diamond", score: 92.6, metric: "accuracy", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "Humanity's Last Exam", score: 43.6, metric: "accuracy", variant: "xhigh, no tools", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "IFBench", score: 82.8, metric: "score", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "OSWorld-Verified", score: 86.1, metric: "success rate", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }, { name: "MMMU Pro", score: 82.3, metric: "accuracy", variant: "xhigh", source: "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421", date: "2026-08-03" }] }, "alibaba/qwen3-coder-30b-a3b-instruct": { id: "alibaba/qwen3-coder-30b-a3b-instruct", name: "Qwen3-Coder 30B-A3B Instruct", description: "Smaller Qwen coder for efficient local agents and repo-level fixes", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-04", last_updated: "2025-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 19.4, metric: "index", source: "https://openrouter.ai/qwen/qwen3-coder-30b-a3b-instruct/benchmarks", date: "2026-06-02" }, { name: "SciCode", score: 27.8, metric: "percent correct", source: "https://openrouter.ai/qwen/qwen3-coder-30b-a3b-instruct/benchmarks", date: "2026-06-02" }, { name: "Terminal-Bench Hard", score: 15.2, metric: "success rate", source: "https://openrouter.ai/qwen/qwen3-coder-30b-a3b-instruct/benchmarks", date: "2026-06-02" }] }, "alibaba/qwen3.7-max": { id: "alibaba/qwen3.7-max", name: "Qwen3.7 Max", description: "Qwen frontier model tuned for agent frameworks, coding assistants, and long tasks", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-05-21", last_updated: "2026-05-21", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 65536 }, benchmarks: [{ name: "SWE-Bench Verified", score: 80.4, metric: "resolved", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "SWE-Bench Pro", score: 60.6, metric: "resolve rate", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "SWE-Bench Multilingual", score: 78.3, metric: "resolve rate", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "Terminal-Bench", score: 69.7, metric: "success rate", harness: "Terminus-2", version: "2.0", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "GPQA Diamond", score: 92.4, metric: "accuracy", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "Humanity's Last Exam", score: 41.4, metric: "accuracy", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "SciCode", score: 53.5, source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "MCP Atlas", score: 76.4, metric: "success rate", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }, { name: "NL2Repo", score: 47.2, harness: "Claude Code", source: "https://qwen.ai/blog?id=qwen3.7", date: "2026-05-19" }] }, "alibaba/qwen-turbo": { id: "alibaba/qwen-turbo", name: "Qwen Turbo", description: "Efficient Qwen model for fast chat, extraction, and high-volume workloads", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2024-11-01", last_updated: "2025-04-28", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 16384 } }, "alibaba/qwen-omni-turbo": { id: "alibaba/qwen-omni-turbo", name: "Qwen-Omni Turbo", description: "Qwen omni model for text, vision, audio, and multimodal agent tasks", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2025-01-19", last_updated: "2025-03-26", modalities: { input: ["text", "image", "audio", "video"], output: ["text", "audio"] }, open_weights: false, limit: { context: 32768, output: 2048 } }, "alibaba/qwen-vl-max": { id: "alibaba/qwen-vl-max", name: "Qwen-VL Max", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2024-04-08", last_updated: "2025-08-13", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 8192 } }, "alibaba/qwen3-32b": { id: "alibaba/qwen3-32b", name: "Qwen3 32B", description: "Dense open Qwen model for self-hosted chat, reasoning, and coding", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-04", last_updated: "2025-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-32B" }], benchmarks: [{ name: "Aider Polyglot", score: 40, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-08" }] }, "alibaba/qwen3-235b-a22b": { id: "alibaba/qwen3-235b-a22b", name: "Qwen3 235B-A22B", description: "Large open Qwen MoE for multilingual reasoning, coding, and tool use", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-04", last_updated: "2025-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-235B-A22B" }], benchmarks: [{ name: "Aider Polyglot", score: 59.6, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-05-09" }, { name: "SWE-Bench Pro", score: 21.41, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "alibaba/qwen-max": { id: "alibaba/qwen-max", name: "Qwen Max", description: "Flagship Qwen model for complex reasoning, coding, and agentic workflows", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2024-04-03", last_updated: "2025-01-25", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 32768, output: 8192 }, benchmarks: [{ name: "Aider Polyglot", score: 21.8, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-01-28" }] }, "alibaba/qwen3.5-35b-a3b": { id: "alibaba/qwen3.5-35b-a3b", name: "Qwen3.5 35B-A3B", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-23", last_updated: "2026-02-23", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.5-35B-A3B" }] }, "alibaba/qwen3-vl-235b-a22b-thinking": { id: "alibaba/qwen3-vl-235b-a22b-thinking", name: "Qwen3 VL 235B A22B Thinking", description: "Qwen vision-language thinking model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-09-23", last_updated: "2025-09-23", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Thinking" }] }, "alibaba/qwen3-30b-a3b": { id: "alibaba/qwen3-30b-a3b", name: "Qwen3 30B A3B", description: "Sparse MoE Qwen model with 3B active parameters for efficient chat and reasoning", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-04-28", last_updated: "2025-04-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-30B-A3B" }] }, "alibaba/qwen3.5-397b-a17b": { id: "alibaba/qwen3.5-397b-a17b", name: "Qwen3.5 397B-A17B", description: "Large open Qwen multimodal MoE for visual agents and long technical tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-15", last_updated: "2026-02-15", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.5-397B-A17B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 76.4, metric: "resolved", source: "https://huggingface.co/Qwen/Qwen3.5-397B-A17B" }] }, "alibaba/qwen3.5-flash": { id: "alibaba/qwen3.5-flash", name: "Qwen3.5 Flash", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-23", last_updated: "2026-02-23", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 65536 } }, "alibaba/qwen2.5-coder-0.5b": { id: "alibaba/qwen2.5-coder-0.5b", name: "Qwen2.5-Coder-0.5B", description: "Tiny open Qwen code model for lightweight completion and on-device coding", family: "qwen", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2024-11-12", last_updated: "2024-11-12", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 32768, output: 8192 }, license: "Apache 2.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B" }] }, "alibaba/qwen-plus": { id: "alibaba/qwen-plus", name: "Qwen Plus", description: "Qwen instruction model for multilingual chat, reasoning, and tool use", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2024-01-25", last_updated: "2025-09-11", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 32768 } }, "alibaba/qwen3-coder-next": { id: "alibaba/qwen3-coder-next", name: "Qwen3 Coder Next", description: "Open-weight Qwen coding model for agents, repository edits, and multi-turn tool use", family: "qwen", attachment: false, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-09", release_date: "2026-02-03", last_updated: "2026-02-03", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-Coder-Next" }] }, "alibaba/qwen3.5-plus": { id: "alibaba/qwen3.5-plus", name: "Qwen3.5 Plus", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2026-02-16", last_updated: "2026-02-16", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 65536 } }, "alibaba/qwen2-5-vl-72b-instruct": { id: "alibaba/qwen2-5-vl-72b-instruct", name: "Qwen2.5-VL 72B Instruct", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2024-09", last_updated: "2024-09", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct" }] }, "alibaba/qwen3-coder-flash": { id: "alibaba/qwen3-coder-flash", name: "Qwen3 Coder Flash", description: "Qwen coding model for software agents, repository edits, and code reasoning", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-07-28", last_updated: "2025-07-28", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 65536 } }, "alibaba/qwen3.6-max-preview": { id: "alibaba/qwen3.6-max-preview", name: "Qwen3.6 Max Preview", description: "Flagship Qwen model for complex reasoning, coding, and agentic workflows", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2026-04-20", last_updated: "2026-04-20", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 262144, output: 65536 } }, "alibaba/qwen3.8-max": { id: "alibaba/qwen3.8-max", name: "Qwen3.8 Max", description: "2.4-trillion-parameter MoE flagship for coding, professional work, multimodal understanding, and long-horizon agentic workflows", family: "qwen", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-08-03", last_updated: "2026-08-03", modalities: { input: ["text", "image", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 131072 } }, "alibaba/qwen3.7-plus": { id: "alibaba/qwen3.7-plus", name: "Qwen3.7 Plus", description: "Multimodal Qwen workhorse for long-context agents, visual inputs, and coding", family: "qwen", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2026-06-02", last_updated: "2026-06-02", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 64e3 } }, "alibaba/qwq-32b": { id: "alibaba/qwq-32b", name: "QwQ 32B", description: "Open reasoning model from the Qwen team for math, coding, and step-by-step problem solving", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2025-03-05", last_updated: "2025-03-05", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/QwQ-32B" }] }, "alibaba/qwen3.7-flash": { id: "alibaba/qwen3.7-flash", name: "Qwen3.7 Flash", description: "Lightweight multimodal Qwen model for high-throughput text, image, and video tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-07-15", last_updated: "2026-07-15", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, input: 991e3, output: 65536 } }, "alibaba/qwen3-next-80b-a3b-thinking": { id: "alibaba/qwen3-next-80b-a3b-thinking", name: "Qwen3-Next 80B-A3B (Thinking)", description: "Efficient Qwen thinking model for local reasoning, math, and coding agents", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-09", last_updated: "2025-09", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking" }] }, "alibaba/qwen3-coder-480b-a35b-instruct": { id: "alibaba/qwen3-coder-480b-a35b-instruct", name: "Qwen3-Coder 480B-A35B Instruct", description: "Open Qwen coding heavyweight for repository reasoning and agentic engineering", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-04", last_updated: "2025-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct" }], benchmarks: [{ name: "SWE-Bench Pro", score: 38.7, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "alibaba/qwen2.5-coder-32b-instruct": { id: "alibaba/qwen2.5-coder-32b-instruct", name: "Qwen2.5-Coder-32B-Instruct", description: "Open coding-focused Qwen model for code generation, repair, and repository reasoning", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2024-11-12", last_updated: "2024-11-12", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct" }] }, "alibaba/qwen-flash": { id: "alibaba/qwen-flash", name: "Qwen Flash", description: "Efficient Qwen model for fast chat, extraction, and high-volume workloads", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2025-07-28", last_updated: "2025-07-28", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 32768 } }, "alibaba/qwen3-max": { id: "alibaba/qwen3-max", name: "Qwen3 Max", description: "Flagship Qwen3 model for coding agents, complex reasoning, and tool use", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-09-23", last_updated: "2025-09-23", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 262144, output: 65536 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 26.4, metric: "index", source: "https://openrouter.ai/qwen/qwen3-max/benchmarks", date: "2026-05-30" }, { name: "SciCode", score: 38.3, metric: "percent correct", source: "https://openrouter.ai/qwen/qwen3-max/benchmarks", date: "2026-05-30" }, { name: "Terminal-Bench Hard", score: 20.5, metric: "success rate", source: "https://openrouter.ai/qwen/qwen3-max/benchmarks", date: "2026-05-30" }] }, "alibaba/qwen3.6-35b-a3b": { id: "alibaba/qwen3.6-35b-a3b", name: "Qwen3.6 35B-A3B", description: "Open multimodal Qwen MoE for local agents that need vision, audio, and code", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-17", last_updated: "2026-04-17", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 73.4, metric: "resolved", source: "https://huggingface.co/Qwen/Qwen3.6-35B-A3B" }] }, "alibaba/qwen3-235b-a22b-instruct-2507": { id: "alibaba/qwen3-235b-a22b-instruct-2507", name: "Qwen3 235B-A22B Instruct 2507", description: "Updated large open Qwen3 MoE instruct model for multilingual chat, coding, and tool use", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2025-07-21", last_updated: "2025-07-21", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 16384 }, license: "Apache 2.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507" }] }, "alibaba/qwen3.6-27b": { id: "alibaba/qwen3.6-27b", name: "Qwen3.6 27B", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-22", last_updated: "2026-04-22", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.6-27B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 77.2, metric: "resolved", source: "https://huggingface.co/Qwen/Qwen3.6-27B" }] }, "alibaba/qwen3-next-80b-a3b-instruct": { id: "alibaba/qwen3-next-80b-a3b-instruct", name: "Qwen3-Next 80B-A3B Instruct", description: "Qwen instruction model for multilingual chat, reasoning, and tool use", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-09", last_updated: "2025-09", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct" }] }, "alibaba/qwen3.8-27b": { id: "alibaba/qwen3.8-27b", name: "Qwen3.8 27B", description: "Dense 27B vision-language model for coding, agent tasks, and image and video understanding", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-14", last_updated: "2026-08-14", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.8-27B" }], benchmarks: [{ name: "SWE-bench Pro", score: 61.7, metric: "resolved", source: "https://huggingface.co/Qwen/Qwen3.8-27B" }] }, "alibaba/qwen3.6-plus": { id: "alibaba/qwen3.6-plus", name: "Qwen3.6 Plus", description: "Earlier Qwen multimodal workhorse for million-token agent and document tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2026-04-02", last_updated: "2026-04-02", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 65536 } }, "alibaba/qwen3.5-9b": { id: "alibaba/qwen3.5-9b", name: "Qwen3.5 9B", description: "Qwen instruction model for multilingual chat, reasoning, and tool use", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-23", last_updated: "2026-02-23", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.5-9B" }] }, "alibaba/qwen3.5-27b": { id: "alibaba/qwen3.5-27b", name: "Qwen3.5 27B", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-23", last_updated: "2026-02-23", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.5-27B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 72.4, metric: "resolved", source: "https://huggingface.co/Qwen/Qwen3.5-27B" }] }, "alibaba/qwq-plus": { id: "alibaba/qwq-plus", name: "QwQ Plus", description: "Qwen reasoning model for deliberate problem solving, math, and coding", family: "qwen", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2025-03-05", last_updated: "2025-03-05", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 8192 } }, "alibaba/qwen3.5-122b-a10b": { id: "alibaba/qwen3.5-122b-a10b", name: "Qwen3.5 122B-A10B", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-02-23", last_updated: "2026-02-23", modalities: { input: ["text", "image", "video", "audio"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 65536 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.5-122B-A10B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 72, metric: "resolved", source: "https://huggingface.co/Qwen/Qwen3.5-122B-A10B" }] }, "alibaba/qwen3-coder-plus": { id: "alibaba/qwen3-coder-plus", name: "Qwen3 Coder Plus", description: "Hosted Qwen coder for software agents, repo edits, and long-context code", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-07-23", last_updated: "2025-07-23", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1048576, output: 65536 } }, "alibaba/qwen3-vl-235b-a22b-instruct": { id: "alibaba/qwen3-vl-235b-a22b-instruct", name: "Qwen3 VL 235B A22B Instruct", description: "Qwen vision-language instruct model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2025-03-31", release_date: "2025-09-23", last_updated: "2025-09-23", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct" }] }, "alibaba/qwen-vl-plus": { id: "alibaba/qwen-vl-plus", name: "Qwen-VL Plus", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-04", release_date: "2024-01-25", last_updated: "2025-08-15", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 8192 } }, "alibaba/qwen3.6-flash": { id: "alibaba/qwen3.6-flash", name: "Qwen3.6 Flash", description: "Qwen vision-language model for visual reasoning, documents, and agent tasks", family: "qwen3.6", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-27", last_updated: "2026-04-27", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 65536 } }, "alibaba/qwen3.8-2.4t-a95b": { id: "alibaba/qwen3.8-2.4t-a95b", name: "Qwen3.8 2.4T A95B", description: "Open-weight sparse MoE (2.4T total, 95B active), the open-weight twin of Qwen3.8 Max for coding, research, complex reasoning, and agentic workflows", family: "qwen", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-12", last_updated: "2026-08-12", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 131072 }, license: "qwen3.8-max", weights: [{ label: "Hugging Face", url: "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B" }] }, "sakana/fugu": { id: "sakana/fugu", name: "Fugu", description: "Multi-agent model for routing expert agents across complex analytical tasks", family: "fugu", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, release_date: "2026-06-15", last_updated: "2026-06-15", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 1e6 }, links: [{ label: "Official model catalog", url: "https://raw.githubusercontent.com/SakanaAI/fugu/refs/heads/main/configs/files/fugu.json", type: "docs" }], benchmarks: [{ name: "SWE Bench Pro", score: 59, source: "https://console.sakana.ai/models" }, { name: "Terminal Bench 2.1", score: 80.2, source: "https://console.sakana.ai/models" }, { name: "LiveCodeBench", score: 92.9, source: "https://console.sakana.ai/models" }, { name: "LiveCodeBench Pro", score: 87.8, source: "https://console.sakana.ai/models" }, { name: "Humanity\u2019s Last Exam", score: 47.2, source: "https://console.sakana.ai/models" }, { name: "CharXiv Reasoning", score: 85.1, source: "https://console.sakana.ai/models" }, { name: "GPQA Diamond", score: 95.5, source: "https://console.sakana.ai/models" }, { name: "SciCode", score: 60.1, source: "https://console.sakana.ai/models" }, { name: "\u03C43 Banking", score: 21.7, source: "https://console.sakana.ai/models" }, { name: "Long Context Reasoning", score: 74.7, source: "https://console.sakana.ai/models" }, { name: "MRCRv2", score: 86.6, source: "https://console.sakana.ai/models" }, { name: "CTI-REALM", score: 67.5, source: "https://console.sakana.ai/models" }] }, "sakana/sakana-namazu": { id: "sakana/sakana-namazu", name: "Sakana Namazu", description: "Japanese-specialized reasoning model based on Kimi K2.6 and tuned for Japanese language, culture, and business workflows", family: "sakana-namazu", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-03", last_updated: "2026-08-03", modalities: { input: ["text", "image", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 262144, output: 65536 }, links: [{ label: "Official product page", url: "https://sakana.ai/namazu/", type: "announcement" }, { label: "Official model documentation", url: "https://console.sakana.ai/models?model=sakana-namazu", type: "docs" }], benchmarks: [{ name: "AIME26", score: 96.67, source: "https://console.sakana.ai/models?model=sakana-namazu" }, { name: "MMLU-Pro", score: 90.33, source: "https://console.sakana.ai/models?model=sakana-namazu" }, { name: "LiveCodeBench v6", score: 90.33, source: "https://console.sakana.ai/models?model=sakana-namazu" }, { name: "JFBench", score: 37.4, source: "https://console.sakana.ai/models?model=sakana-namazu" }, { name: "Translation", score: 52.2, source: "https://console.sakana.ai/models?model=sakana-namazu" }, { name: "FairPoliticsQA", score: 56.3, source: "https://console.sakana.ai/models?model=sakana-namazu" }] }, "sakana/fugu-ultra": { id: "sakana/fugu-ultra", name: "Fugu Ultra", description: "Quality-first multi-agent model for hard research, analysis, and competitions", family: "fugu", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: false, release_date: "2026-06-15", last_updated: "2026-06-15", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 1e6 }, links: [{ label: "Official model catalog", url: "https://raw.githubusercontent.com/SakanaAI/fugu/refs/heads/main/configs/files/fugu.json", type: "docs" }], benchmarks: [{ name: "SWE Bench Pro", score: 73.7, source: "https://console.sakana.ai/models" }, { name: "Terminal Bench 2.1", score: 82.1, source: "https://console.sakana.ai/models" }, { name: "LiveCodeBench", score: 93.2, source: "https://console.sakana.ai/models" }, { name: "LiveCodeBench Pro", score: 90.8, source: "https://console.sakana.ai/models" }, { name: "Humanity\u2019s Last Exam", score: 50, source: "https://console.sakana.ai/models" }, { name: "CharXiv Reasoning", score: 86.6, source: "https://console.sakana.ai/models" }, { name: "GPQA Diamond", score: 95.5, source: "https://console.sakana.ai/models" }, { name: "SciCode", score: 58.7, source: "https://console.sakana.ai/models" }, { name: "\u03C43 Banking", score: 20.6, source: "https://console.sakana.ai/models" }, { name: "Long Context Reasoning", score: 73.3, source: "https://console.sakana.ai/models" }, { name: "MRCRv2", score: 93.6, source: "https://console.sakana.ai/models" }, { name: "CTI-REALM", score: 69.4, source: "https://console.sakana.ai/models" }] }, "perplexity/sonar-deep-research": { id: "perplexity/sonar-deep-research", name: "Sonar Deep Research", description: "Sonar search model for autonomous research and citation-backed long-form reports", family: "sonar", attachment: false, reasoning: true, tool_call: false, temperature: false, knowledge: "2025-01", release_date: "2025-02-01", last_updated: "2025-09-01", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 32768 } }, "perplexity/sonar": { id: "perplexity/sonar", name: "Sonar", description: "Fast web-grounded Sonar for current answers, citations, and lightweight retrieval", family: "sonar", attachment: false, reasoning: false, tool_call: false, temperature: true, knowledge: "2025-09-01", release_date: "2024-01-01", last_updated: "2025-09-01", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 4096 }, benchmarks: [{ name: "SciCode", score: 22.9, metric: "percent correct", source: "https://openrouter.ai/perplexity/sonar/benchmarks", date: "2026-03-11" }] }, "perplexity/sonar-reasoning-pro": { id: "perplexity/sonar-reasoning-pro", name: "Sonar Reasoning Pro", description: "Web-grounded Sonar for multi-step research questions that need cited reasoning", family: "sonar-reasoning", attachment: true, reasoning: true, tool_call: false, temperature: true, knowledge: "2025-09-01", release_date: "2024-01-01", last_updated: "2025-09-01", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 4096 } }, "perplexity/sonar-pro": { id: "perplexity/sonar-pro", name: "Sonar Pro", description: "Deeper Sonar search model with broader retrieval and stronger synthesis", family: "sonar-pro", attachment: true, reasoning: false, tool_call: false, temperature: true, knowledge: "2025-09-01", release_date: "2024-01-01", last_updated: "2025-09-01", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 8192 }, benchmarks: [{ name: "SciCode", score: 22.6, metric: "percent correct", source: "https://openrouter.ai/perplexity/sonar-pro/benchmarks", date: "2026-03-11" }] }, "zhipuai/glm-4.6": { id: "zhipuai/glm-4.6", name: "GLM-4.6", description: "Late GLM-4 workhorse for coding agents, reasoning, and structured tasks", family: "glm", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-09-30", last_updated: "2025-09-30", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.6" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 29.5, metric: "index", source: "https://openrouter.ai/z-ai/glm-4.6/benchmarks", date: "2026-05-22" }, { name: "SciCode", score: 38.4, metric: "percent correct", source: "https://openrouter.ai/z-ai/glm-4.6/benchmarks", date: "2026-05-22" }, { name: "Terminal-Bench Hard", score: 25, metric: "success rate", source: "https://openrouter.ai/z-ai/glm-4.6/benchmarks", date: "2026-05-22" }, { name: "SWE-Bench Pro", score: 9.67, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "zhipuai/glm-4.5-flash": { id: "zhipuai/glm-4.5-flash", name: "GLM-4.5-Flash", description: "Efficient GLM model for fast reasoning, coding, and agent workflows", family: "glm-flash", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-07-28", last_updated: "2025-07-28", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 98304 } }, "zhipuai/glm-5v-turbo": { id: "zhipuai/glm-5v-turbo", name: "GLM-5V-Turbo", description: "Fast GLM vision model for screenshots, documents, and multimodal agent tasks", family: "glm", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-04-01", last_updated: "2026-04-01", modalities: { input: ["text", "image", "video", "pdf"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 131072 } }, "zhipuai/glm-4.7": { id: "zhipuai/glm-4.7", name: "GLM-4.7", description: "Mature GLM model for dependable coding, reasoning, and structured agent tasks", family: "glm", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-12-22", last_updated: "2025-12-22", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.7" }], benchmarks: [{ name: "SWE-Bench Verified", score: 73.8, metric: "resolved", source: "https://huggingface.co/zai-org/GLM-4.7" }, { name: "Terminal Bench 2.0", score: 33.4, metric: "score", source: "https://huggingface.co/zai-org/GLM-4.7" }] }, "zhipuai/glm-4.5v": { id: "zhipuai/glm-4.5v", name: "GLM-4.5V", description: "GLM vision model for visual reasoning, documents, and multimodal agents", family: "glm", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-08-11", last_updated: "2025-08-11", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 64e3, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.5V" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 10.9, metric: "index", source: "https://openrouter.ai/z-ai/glm-4.5v/benchmarks", date: "2026-04-29" }, { name: "SciCode", score: 22.1, metric: "percent correct", source: "https://openrouter.ai/z-ai/glm-4.5v/benchmarks", date: "2026-04-29" }, { name: "Terminal-Bench Hard", score: 5.3, metric: "success rate", source: "https://openrouter.ai/z-ai/glm-4.5v/benchmarks", date: "2026-04-29" }] }, "zhipuai/glm-4.6v-flash": { id: "zhipuai/glm-4.6v-flash", name: "GLM-4.6V-Flash", description: "Lightweight GLM vision model for visual reasoning, documents, and multimodal agents", family: "glm", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2025-12-08", last_updated: "2025-12-08", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.6V-Flash" }] }, "zhipuai/glm-4.7-flashx": { id: "zhipuai/glm-4.7-flashx", name: "GLM-4.7-FlashX", description: "Efficient GLM model for fast reasoning, coding, and agent workflows", family: "glm-flash", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2026-01-19", last_updated: "2026-01-19", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 2e5, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.7-Flash" }] }, "zhipuai/glm-5": { id: "zhipuai/glm-5", name: "GLM-5", description: "General GLM flagship for coding, analysis, and tool-heavy engineering workflows", family: "glm", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-12", last_updated: "2026-02-12", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-5" }], benchmarks: [{ name: "SWE-Bench Verified", score: 72.8, metric: "resolved", source: "https://www.swebench.com/" }, { name: "SWE-Atlas Codebase QnA", score: 20.5, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 24.24, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 28.74, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }] }, "zhipuai/glm-5-turbo": { id: "zhipuai/glm-5-turbo", name: "GLM-5-Turbo", description: "Faster GLM-5 lane for coding agents that need lower latency", family: "glm", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-03-16", last_updated: "2026-03-16", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 131072 } }, "zhipuai/glm-5.3": { id: "zhipuai/glm-5.3", name: "GLM-5.3", description: "Flagship GLM model for long-horizon coding, agents, and complex project delivery", family: "glm", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-08-14", last_updated: "2026-08-14", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 1e6, output: 131072 } }, "zhipuai/glm-5.2": { id: "zhipuai/glm-5.2", name: "GLM-5.2", description: "Open flagship GLM for long-horizon coding agents and million-token context work", family: "glm", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-06-13", last_updated: "2026-06-13", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 1e6, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-5.2" }], benchmarks: [{ name: "SWE-Bench Pro", score: 62.1, metric: "resolve rate", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "Terminal-Bench", score: 82.7, metric: "success rate", harness: "Claude Code", version: "2.1", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "FrontierSWE", score: 74.4, metric: "dominance", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "Humanity's Last Exam", score: 40.5, metric: "accuracy", dataset: "text-only subset", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "Humanity's Last Exam", score: 54.7, metric: "accuracy", variant: "with tools", dataset: "text-only subset", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "CritPt", score: 20.9, metric: "accuracy", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "AIME", score: 99.2, metric: "accuracy", version: "2026", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "HMMT", score: 94.4, metric: "accuracy", version: "November 2025", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "HMMT", score: 92.5, metric: "accuracy", version: "February 2026", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "IMOAnswerBench", score: 91, metric: "accuracy", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "GPQA Diamond", score: 91.2, metric: "accuracy", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "NL2Repo", score: 48.9, metric: "resolve rate", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "DeepSWE", score: 46.2, metric: "resolve rate", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "Program Bench", score: 63.7, metric: "score", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "Terminal-Bench", score: 81, metric: "success rate", harness: "Terminus 2", version: "2.1", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "PostTrainBench", score: 34.3, metric: "score", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "SWE Marathon", score: 13, metric: "resolve rate", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "MCP Atlas", score: 76.8, metric: "score", dataset: "public subset", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }, { name: "Tool-Decathlon", score: 48.2, metric: "score", source: "https://z.ai/blog/glm-5.2", date: "2026-06-16" }] }, "zhipuai/glm-4.6v": { id: "zhipuai/glm-4.6v", name: "GLM-4.6V", description: "GLM vision model for visual reasoning, documents, and multimodal agents", family: "glm", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-12-08", last_updated: "2025-12-08", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 32768 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.6V" }] }, "zhipuai/glm-4.5-air": { id: "zhipuai/glm-4.5-air", name: "GLM-4.5-Air", description: "Lighter GLM-4.5 variant for fast coding assistance and cheaper agents", family: "glm-air", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-07-28", last_updated: "2025-07-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 98304 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.5-Air" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 23.8, metric: "index", source: "https://openrouter.ai/z-ai/glm-4.5-air/benchmarks", date: "2026-05-30" }, { name: "SciCode", score: 30.6, metric: "percent correct", source: "https://openrouter.ai/z-ai/glm-4.5-air/benchmarks", date: "2026-05-30" }, { name: "Terminal-Bench Hard", score: 20.5, metric: "success rate", source: "https://openrouter.ai/z-ai/glm-4.5-air/benchmarks", date: "2026-05-30" }] }, "zhipuai/glm-5.1": { id: "zhipuai/glm-5.1", name: "GLM-5.1", description: "Strong GLM coding model for agentic engineering, terminals, and repository generation", family: "glm", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-07", last_updated: "2026-04-07", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 2e5, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-5.1" }], benchmarks: [{ name: "Artificial Analysis Coding Agent Index", score: 52.7, metric: "average pass@1", harness: "Claude Code", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Atlas Codebase QnA", score: 73.2, metric: "pass@1", harness: "Claude Code", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "SWE-Bench Pro", score: 19.8, metric: "pass@1", harness: "Claude Code", dataset: "hard-aa", source: "https://artificialanalysis.ai/agents/coding-agents" }, { name: "Terminal-Bench", score: 65.1, metric: "pass@1", harness: "Claude Code", version: "2.1", source: "https://artificialanalysis.ai/agents/coding-agents" }] }, "zhipuai/glm-4.7-flash": { id: "zhipuai/glm-4.7-flash", name: "GLM-4.7-Flash", description: "Budget GLM lane for fast coding help, routing, and everyday automation", family: "glm-flash", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2026-01-19", last_updated: "2026-01-19", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 2e5, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.7-Flash" }], benchmarks: [{ name: "SWE-Bench Verified", score: 59.2, metric: "resolved", source: "https://huggingface.co/zai-org/GLM-4.7-Flash" }] }, "zhipuai/glm-4.5": { id: "zhipuai/glm-4.5", name: "GLM-4.5", description: "Hybrid-reasoning GLM release that made the 4.5 line broadly useful", family: "glm", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-04", release_date: "2025-07-28", last_updated: "2025-07-28", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 98304 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/zai-org/GLM-4.5" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 26.3, metric: "index", source: "https://openrouter.ai/z-ai/glm-4.5/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 34.8, metric: "percent correct", source: "https://openrouter.ai/z-ai/glm-4.5/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 22, metric: "success rate", source: "https://openrouter.ai/z-ai/glm-4.5/benchmarks", date: "2026-03-11" }] }, "ibm/granite-4-h-micro": { id: "ibm/granite-4-h-micro", name: "Granite-4.0-H-Micro", description: "Compact open-weight hybrid Granite model for lightweight enterprise chat and tool calling", family: "granite", attachment: false, reasoning: false, tool_call: true, structured_output: true, temperature: true, release_date: "2025-10-02", last_updated: "2025-10-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/ibm-granite/granite-4.0-h-micro" }] }, "ibm/granite-4-h-small": { id: "ibm/granite-4-h-small", name: "Granite-4.0-H-Small", description: "Open-weight hybrid model for enterprise chat, coding, retrieval-augmented generation, and tool-calling workloads", family: "granite", attachment: false, reasoning: false, tool_call: true, structured_output: true, temperature: true, release_date: "2025-10-02", last_updated: "2025-10-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/ibm-granite/granite-4.0-h-small" }] }, "thinkingmachines/inkling-small": { id: "thinkingmachines/inkling-small", name: "Inkling Small", description: "Multimodal MoE reasoning model (276B total, 12B active) for text, image, and audio", family: "ling", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-07-30", last_updated: "2026-07-30", modalities: { input: ["text", "image", "audio"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 1048576 }, license: "Apache-2.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/thinkingmachines/Inkling-Small" }] }, "thinkingmachines/inkling": { id: "thinkingmachines/inkling", name: "Inkling", description: "Multimodal MoE reasoning model (975B total, 41B active) for text, image, and audio", family: "ling", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-07-15", last_updated: "2026-07-15", modalities: { input: ["text", "image", "audio"], output: ["text"] }, open_weights: true, limit: { context: 1048576, output: 1048576 }, license: "Apache-2.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/thinkingmachines/Inkling" }] }, "mistral/magistral-small-2506": { id: "mistral/magistral-small-2506", name: "Magistral Small", description: "Open Mistral reasoning model for transparent step-by-step problem solving", family: "magistral", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-06-10", last_updated: "2025-06-10", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, license: "Apache 2.0", weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Magistral-Small-2506" }] }, "mistral/mistral-small-latest": { id: "mistral/mistral-small-latest", name: "Mistral Small (latest)", description: "Efficient Mistral model for fast chat, extraction, and production assistants", family: "mistral-small", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-06", release_date: "2026-03-16", last_updated: "2026-03-16", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 256e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }] }, "mistral/devstral-medium-2507": { id: "mistral/devstral-medium-2507", name: "Devstral Medium", description: "Mistral coding agent model for repository tasks and software engineering workflows", family: "devstral", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-05", release_date: "2025-07-10", last_updated: "2025-07-10", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 128e3 }, benchmarks: [{ name: "SWE-Bench Verified", score: 61.6, metric: "resolved", source: "https://mistral.ai/news/devstral-2507", date: "2025-07-10" }] }, "mistral/mistral-small-3-1-24b-instruct-2503": { id: "mistral/mistral-small-3-1-24b-instruct-2503", name: "Mistral Small 3.1 24B", description: "Efficient multimodal model for instruction following, coding, reasoning, and function calling", family: "mistral-small", attachment: true, reasoning: false, tool_call: true, structured_output: true, temperature: true, knowledge: "2024-06", release_date: "2025-03-17", last_updated: "2025-03-17", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503" }] }, "mistral/mistral-large-2411": { id: "mistral/mistral-large-2411", name: "Mistral Large 2.1", description: "Flagship Mistral model for advanced reasoning, coding, and multilingual work", family: "mistral-large", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-11", release_date: "2024-11-18", last_updated: "2024-11-18", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 13.8, metric: "index", source: "https://openrouter.ai/mistralai/mistral-large-2407/benchmarks", date: "2026-03-11" }, { name: "SciCode", score: 29.2, metric: "percent correct", source: "https://openrouter.ai/mistralai/mistral-large-2407/benchmarks", date: "2026-03-11" }, { name: "Terminal-Bench Hard", score: 6.1, metric: "success rate", source: "https://openrouter.ai/mistralai/mistral-large-2407/benchmarks", date: "2026-03-11" }] }, "mistral/mistral-nemo": { id: "mistral/mistral-nemo", name: "Mistral Nemo", description: "Efficient Mistral-NVIDIA open model for multilingual chat and local deployment", family: "mistral-nemo", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-07", release_date: "2024-07-01", last_updated: "2024-07-01", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407" }] }, "mistral/mistral-large-latest": { id: "mistral/mistral-large-latest", name: "Mistral Large (latest)", description: "Flagship Mistral model for advanced reasoning, coding, and multilingual work", family: "mistral-large", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-11", release_date: "2024-11-01", last_updated: "2025-12-02", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512" }] }, "mistral/ministral-8b-instruct-2410": { id: "mistral/ministral-8b-instruct-2410", name: "Ministral 8B Instruct", description: "Efficient open Mistral edge model for on-device chat and function calling", family: "ministral", attachment: false, reasoning: false, tool_call: true, temperature: true, release_date: "2024-10-16", last_updated: "2024-10-16", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 131072, output: 8192 }, license: "Mistral Research License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Ministral-8B-Instruct-2410" }] }, "mistral/devstral-small-2507": { id: "mistral/devstral-small-2507", name: "Devstral Small", description: "Mistral coding agent model for repository tasks and software engineering workflows", family: "devstral", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-05", release_date: "2025-07-10", last_updated: "2025-07-10", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Devstral-Small-2507" }], benchmarks: [{ name: "SWE-Bench Verified", score: 53.6, metric: "resolved", source: "https://mistral.ai/news/devstral-2507", date: "2025-07-10" }] }, "mistral/devstral-2512": { id: "mistral/devstral-2512", name: "Devstral 2", description: "Mistral's coding-agent model for repository work, terminal tasks, and software fixes", family: "devstral", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-12", release_date: "2025-12-09", last_updated: "2025-12-09", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 23.7, metric: "index", source: "https://openrouter.ai/mistralai/devstral-2512/benchmarks", date: "2026-05-31" }, { name: "SciCode", score: 33.1, metric: "percent correct", source: "https://openrouter.ai/mistralai/devstral-2512/benchmarks", date: "2026-05-31" }, { name: "Terminal-Bench Hard", score: 18.9, metric: "success rate", source: "https://openrouter.ai/mistralai/devstral-2512/benchmarks", date: "2026-05-31" }] }, "mistral/mistral-small-2603": { id: "mistral/mistral-small-2603", name: "Mistral Small 4", description: "Fast Mistral production model for chat, extraction, and cost-sensitive agents", family: "mistral-small", attachment: true, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-06", release_date: "2026-03-16", last_updated: "2026-03-16", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 256e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Small-4-119B-2603" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 24.3, metric: "index", source: "https://openrouter.ai/mistralai/mistral-small-2603/benchmarks", date: "2026-06-01" }, { name: "SciCode", score: 38, metric: "percent correct", source: "https://openrouter.ai/mistralai/mistral-small-2603/benchmarks", date: "2026-06-01" }, { name: "Terminal-Bench Hard", score: 17.4, metric: "success rate", source: "https://openrouter.ai/mistralai/mistral-small-2603/benchmarks", date: "2026-06-01" }] }, "mistral/mistral-small-2506": { id: "mistral/mistral-small-2506", name: "Mistral Small 3.2", description: "Efficient Mistral model for fast chat, extraction, and production assistants", family: "mistral-small", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-03", release_date: "2025-06-20", last_updated: "2025-06-20", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 16384 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506" }] }, "mistral/pixtral-large-latest": { id: "mistral/pixtral-large-latest", name: "Pixtral Large (latest)", description: "Mistral's larger vision model for document-heavy image understanding and chat", family: "pixtral", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-11", release_date: "2024-11-01", last_updated: "2024-11-04", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Pixtral-Large-Instruct-2411" }] }, "mistral/mistral-large-2512": { id: "mistral/mistral-large-2512", name: "Mistral Large 3", description: "Mistral's largest general model for enterprise agents, coding, and multilingual reasoning", family: "mistral-large", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-11", release_date: "2024-11-01", last_updated: "2025-12-02", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Large-3-675B-Instruct-2512" }], benchmarks: [{ name: "Artificial Analysis Coding Index", score: 22.7, metric: "index", source: "https://openrouter.ai/mistralai/mistral-large-2512/benchmarks", date: "2026-06-04" }, { name: "SciCode", score: 36.2, metric: "percent correct", source: "https://openrouter.ai/mistralai/mistral-large-2512/benchmarks", date: "2026-06-04" }, { name: "Terminal-Bench Hard", score: 15.9, metric: "success rate", source: "https://openrouter.ai/mistralai/mistral-large-2512/benchmarks", date: "2026-06-04" }] }, "mistral/mistral-medium-2604": { id: "mistral/mistral-medium-2604", name: "Mistral Medium 3.5", description: "Balanced Mistral model for enterprise assistants, multilingual work, and tools", family: "mistral-medium", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-29", last_updated: "2026-04-29", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 77.6, metric: "resolved", source: "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B" }, { name: "\u03C4\xB3-Telecom", score: 91.4, metric: "accuracy", variant: "public preview", source: "https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/", date: "2026-05-22" }] }, "mistral/mistral-medium-2505": { id: "mistral/mistral-medium-2505", name: "Mistral Medium 3", description: "Mistral model for multilingual chat, reasoning, and tool-assisted workflows", family: "mistral-medium", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-05", release_date: "2025-05-07", last_updated: "2025-05-07", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 131072 }, benchmarks: [{ name: "Artificial Analysis Coding Index", score: 13.6, metric: "index", source: "https://openrouter.ai/mistralai/mistral-medium-3/benchmarks", date: "2026-05-30" }, { name: "SciCode", score: 33.1, metric: "percent correct", source: "https://openrouter.ai/mistralai/mistral-medium-3/benchmarks", date: "2026-05-30" }, { name: "Terminal-Bench Hard", score: 3.8, metric: "success rate", source: "https://openrouter.ai/mistralai/mistral-medium-3/benchmarks", date: "2026-05-30" }] }, "mistral/devstral-medium-latest": { id: "mistral/devstral-medium-latest", name: "Devstral 2 (latest)", description: "Mistral coding agent model for repository tasks and software engineering workflows", family: "devstral", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2025-12", release_date: "2025-12-02", last_updated: "2025-12-02", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512" }] }, "mistral/pixtral-12b": { id: "mistral/pixtral-12b", name: "Pixtral 12B", description: "Mistral vision-language model for image understanding and multimodal chat", family: "pixtral", attachment: true, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-09", release_date: "2024-09-01", last_updated: "2024-09-01", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 128e3, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Pixtral-12B-2409" }] }, "mistral/voxtral-small-latest": { id: "mistral/voxtral-small-latest", name: "Voxtral Small (latest)", description: "Instruct model with native audio input for speech understanding and tool use", family: "voxtral", attachment: true, reasoning: false, tool_call: true, temperature: true, release_date: "2025-07-15", last_updated: "2025-07-15", modalities: { input: ["text", "audio"], output: ["text"] }, open_weights: true, limit: { context: 32e3, output: 32e3 } }, "mistral/codestral-22b-v0.1": { id: "mistral/codestral-22b-v0.1", name: "Codestral-22B-v0.1", description: "Open Mistral code model for fill-in-the-middle and 80+ programming languages", family: "codestral", attachment: false, reasoning: false, tool_call: false, temperature: true, release_date: "2024-05-29", last_updated: "2024-05-29", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 32768, output: 8192 }, license: "Mistral AI Non-Production License", weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Codestral-22B-v0.1" }] }, "mistral/codestral-latest": { id: "mistral/codestral-latest", name: "Codestral (latest)", description: "Mistral code model for completions, refactors, and developer IDE workflows", family: "codestral", attachment: false, reasoning: false, tool_call: true, temperature: true, knowledge: "2024-10", release_date: "2024-05-29", last_updated: "2025-01-04", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 256e3, output: 4096 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Codestral-22B-v0.1" }], benchmarks: [{ name: "Aider Polyglot", score: 11.1, metric: "percent correct", source: "https://aider.chat/docs/leaderboards/", date: "2025-01-13" }] }, "mistral/magistral-medium-latest": { id: "mistral/magistral-medium-latest", name: "Magistral Medium (latest)", description: "Mistral reasoning model for transparent analysis, math, and complex decisions", family: "magistral-medium", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-06", release_date: "2025-03-17", last_updated: "2025-03-20", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 128e3, output: 16384 } }, "mistral/mistral-medium-latest": { id: "mistral/mistral-medium-latest", name: "Mistral Medium (latest)", description: "Balanced Mistral model for enterprise assistants, multilingual work, and tools", family: "mistral-medium", attachment: true, reasoning: true, tool_call: true, structured_output: true, temperature: true, release_date: "2026-04-29", last_updated: "2026-04-29", modalities: { input: ["text", "image"], output: ["text"] }, open_weights: true, limit: { context: 262144, output: 262144 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B" }], benchmarks: [{ name: "SWE-Bench Verified", score: 77.6, metric: "resolved", source: "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B" }] }, "upstage/solar-pro2": { id: "upstage/solar-pro2", name: "Solar Pro 2", description: "Flagship model for demanding analysis, coding, and production agent workflows", family: "solar-pro", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03", release_date: "2025-05-20", last_updated: "2025-05-20", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 65536, output: 8192 } }, "upstage/solar-pro4": { id: "upstage/solar-pro4", name: "Solar Pro 4", description: "Upstage's flagship model, specialized for agentic use", family: "solar-pro", attachment: false, reasoning: true, tool_call: true, structured_output: true, temperature: true, knowledge: "2026-02", release_date: "2026-08-06", last_updated: "2026-08-06", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 524288, output: 131072 } }, "upstage/solar-pro3": { id: "upstage/solar-pro3", name: "Solar Pro 3", description: "Flagship model for demanding analysis, coding, and production agent workflows", family: "solar-pro", attachment: false, reasoning: true, tool_call: true, temperature: true, knowledge: "2025-03", release_date: "2026-01", last_updated: "2026-01", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 131072, output: 8192 } }, "minimax/MiniMax-M2.7": { id: "minimax/MiniMax-M2.7", name: "MiniMax-M2.7", description: "Open MiniMax flagship for coding agents, office automation, and complex environments", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-03-18", last_updated: "2026-03-18", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M2.7" }], benchmarks: [{ name: "SWE-Bench Verified", score: 79.9, metric: "resolved", harness: "Claude Code", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "SWE-Bench Pro", score: 56.2, metric: "resolve rate", harness: "Claude Code", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "Terminal-Bench", score: 51.1, metric: "success rate", version: "2.1", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }] }, "minimax/MiniMax-M2.7-highspeed": { id: "minimax/MiniMax-M2.7-highspeed", name: "MiniMax-M2.7-highspeed", description: "Low-latency M2.7 variant for interactive coding plans and agent loops", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-03-18", last_updated: "2026-03-18", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M2.7" }] }, "minimax/MiniMax-M2.1": { id: "minimax/MiniMax-M2.1", name: "MiniMax-M2.1", description: "Earlier MiniMax agent model for practical coding and productivity tasks", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-12-23", last_updated: "2025-12-23", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M2.1" }], benchmarks: [{ name: "SWE-Bench Verified", score: 74, metric: "resolved", source: "https://huggingface.co/MiniMaxAI/MiniMax-M2.1" }, { name: "SWE-Bench Pro", score: 36.81, metric: "resolve rate", dataset: "public", source: "https://labs.scale.com/leaderboard/swe_bench_pro_public" }] }, "minimax/MiniMax-M2": { id: "minimax/MiniMax-M2", name: "MiniMax-M2", description: "Efficient open MiniMax model built for coding agents and tool-heavy workflows", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2025-10-27", last_updated: "2025-10-27", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 196608, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M2" }], benchmarks: [{ name: "SWE-Bench Verified", score: 69.4, metric: "resolved", source: "https://huggingface.co/MiniMaxAI/MiniMax-M2" }] }, "minimax/MiniMax-M2.5-highspeed": { id: "minimax/MiniMax-M2.5-highspeed", name: "MiniMax-M2.5-highspeed", description: "High-speed MiniMax model for low-latency coding and agent workflows", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-13", last_updated: "2026-02-13", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M2.5" }] }, "minimax/MiniMax-M3": { id: "minimax/MiniMax-M3", name: "MiniMax-M3", description: "MiniMax multimodal model for long-context coding, perception, and agent planning", family: "minimax", attachment: true, reasoning: true, tool_call: true, temperature: true, release_date: "2026-06-01", last_updated: "2026-06-01", modalities: { input: ["text", "image", "video"], output: ["text"] }, open_weights: true, limit: { context: 512e3, output: 128e3 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M3" }], benchmarks: [{ name: "SWE-Bench Verified", score: 80.5, metric: "resolved", harness: "Claude Code", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "SWE-Bench Pro", score: 59, metric: "resolve rate", harness: "Claude Code", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "Terminal-Bench", score: 66, metric: "success rate", version: "2.1", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "BrowseComp", score: 83.52, metric: "accuracy", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "MCP Atlas", score: 74.2, metric: "success rate", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }, { name: "OSWorld-Verified", score: 70.06, metric: "success rate", source: "https://www.minimax.io/blog/minimax-m3", date: "2026-06-01" }] }, "minimax/MiniMax-M2-Her": { id: "minimax/MiniMax-M2-Her", name: "MiniMax-M2 Her", description: "MiniMax M2 variant tuned for conversational and character-driven agent interactions", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-01-23", last_updated: "2026-01-23", modalities: { input: ["text"], output: ["text"] }, open_weights: false, limit: { context: 2e5, output: 131e3 } }, "minimax/MiniMax-M2.5": { id: "minimax/MiniMax-M2.5", name: "MiniMax-M2.5", description: "Prior MiniMax coding model for agent workflows, office edits, and automation", family: "minimax", attachment: false, reasoning: true, tool_call: true, temperature: true, release_date: "2026-02-12", last_updated: "2026-02-12", modalities: { input: ["text"], output: ["text"] }, open_weights: true, limit: { context: 204800, output: 131072 }, weights: [{ label: "Hugging Face", url: "https://huggingface.co/MiniMaxAI/MiniMax-M2.5" }], benchmarks: [{ name: "SWE-Bench Verified", score: 75.8, metric: "resolved", source: "https://www.swebench.com/" }, { name: "SWE-Atlas Codebase QnA", score: 10.3, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-qna" }, { name: "SWE-Atlas Refactoring", score: 19.52, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-refactoring" }, { name: "SWE-Atlas Test Writing", score: 18.6, metric: "score", harness: "Mini-SWE-Agent", source: "https://labs.scale.com/leaderboard/sweatlas-tw" }] } } };
 
@@ -48283,7 +48406,7 @@ function diskCacheFresh() {
   }
 }
 async function fetchFresh() {
-  const dispatcher = proxyDispatcher(process.env.https_proxy || process.env.HTTPS_PROXY || process.env.http_proxy || process.env.HTTP_PROXY);
+  const dispatcher = proxyDispatcher(process.env.https_proxy || process.env.HTTPS_PROXY || process.env.http_proxy || process.env.HTTP_PROXY, 15e3);
   const attempts = dispatcher ? [{ opts: { dispatcher }, label: "via proxy" }, { opts: {}, label: "direct" }] : [{ opts: {}, label: "direct" }];
   for (const { opts, label } of attempts) {
     try {
@@ -49950,8 +50073,17 @@ var SessionStore = class {
   async loadAll() {
     const out = /* @__PURE__ */ new Map();
     if (!this.enabled) return out;
+    let clamped = 0;
     for (const [id, envelope] of await this.store.loadAll()) {
-      out.set(id, buildSession(envelope.payload));
+      const session = buildSession(envelope.payload);
+      if (hasNegativePersistedTokens(envelope.payload)) {
+        await this.store.writeNow(id, () => buildRecord(session));
+        clamped++;
+      }
+      out.set(id, session);
+    }
+    if (clamped > 0) {
+      log("info", `[persist] one-time migration (#408): clamped negative lastInputTokens/contextTokens in ${clamped} session(s) to 0`);
     }
     return out;
   }
@@ -50073,7 +50205,14 @@ var SessionStore = class {
       meta?.protocol ? this.store.loadSync(id, relPathFor(id)) : null
     ];
     for (const envelope of envelopes) {
-      if (envelope) return buildSession(envelope.payload);
+      if (envelope) {
+        const session = buildSession(envelope.payload);
+        if (hasNegativePersistedTokens(envelope.payload)) {
+          this.scheduleSave(session);
+          log("info", `[persist] clamped negative token stats on reload for ${id} (#408)`);
+        }
+        return session;
+      }
     }
     return null;
   }
@@ -50115,13 +50254,15 @@ var SessionStore = class {
   }
 };
 function buildRecord(session) {
+  const snapshot = boundedFoldedSnapshot(session);
   return {
     version: PERSIST_VERSION,
     savedAt: Date.now(),
     id: session.id,
     meta: { ...session.meta },
     stats: { ...session.stats },
-    messages: session.lastMessages,
+    messages: snapshot,
+    messagesFolded: snapshot ? true : void 0,
     metadata: { ...session.metadata },
     state: session.state,
     blockContents: Object.fromEntries(session.blockContents),
@@ -50161,10 +50302,14 @@ function buildSession(parsed) {
       cachedTokens: stats.cachedTokens ?? parsed.cachedTokens ?? 0,
       outputTokens: stats.outputTokens ?? parsed.outputTokens ?? 0,
       cacheSamples: stats.cacheSamples ?? parsed.cacheSamples ?? 0,
-      lastInputTokens: stats.lastInputTokens ?? parsed.lastInputTokens ?? 0,
+      // #408: clamp at restore — pre-clamp versions persisted negative
+      // values (lastInputTokens = total − credit before the Math.max
+      // guard existed) which would otherwise revive after upgrade and
+      // feed the /acp panel + web stats as negative percentages.
+      lastInputTokens: Math.max(0, stats.lastInputTokens ?? parsed.lastInputTokens ?? 0),
       // In-memory only — a fresh process has no pending compress fold.
       compressCreditTokens: 0,
-      contextTokens: stats.contextTokens ?? parsed.contextTokens ?? 0
+      contextTokens: Math.max(0, stats.contextTokens ?? parsed.contextTokens ?? 0)
     },
     metadata: parsed.metadata ?? {},
     state: mergeState(parsed.state),
@@ -50178,6 +50323,7 @@ function buildSession(parsed) {
     restored: true,
     blockContents,
     lastMessages: Array.isArray(parsed.messages) ? parsed.messages : void 0,
+    lastMessagesFolded: parsed.messagesFolded === true,
     inFlight: 0,
     persisted: true
   };
@@ -50187,6 +50333,12 @@ function isValidRecord(parsed) {
   const r = parsed;
   return typeof r.id === "string" && typeof r.state === "object" && r.state !== null && Array.isArray(r.state.blocks);
 }
+function hasNegativePersistedTokens(parsed) {
+  const stats = parsed.stats ?? {};
+  const last = stats.lastInputTokens ?? parsed.lastInputTokens;
+  const ctx = stats.contextTokens ?? parsed.contextTokens;
+  return typeof last === "number" && last < 0 || typeof ctx === "number" && ctx < 0;
+}
 function defaultDir() {
   return sessionsDir();
 }
@@ -50203,6 +50355,36 @@ function persistEnabled() {
   if (env === "0" || env === "false") return false;
   return true;
 }
+function persistTailTokens() {
+  const env = process.env.BILI_PERSIST_TAIL_TOKENS;
+  if (env) {
+    const n = Number.parseInt(env, 10);
+    if (Number.isFinite(n) && n >= 0) return n;
+  }
+  return 16384;
+}
+function boundedFoldedSnapshot(session) {
+  const msgs = session.lastMessages;
+  if (!msgs || msgs.length === 0) return void 0;
+  const budget = persistTailTokens();
+  if (budget === 0) return void 0;
+  let view = prune(msgs, session.state);
+  let total = 0;
+  for (const m2 of view) total += defaultCountTokens(m2.text ?? "");
+  if (total > budget) {
+    let acc = 0;
+    let start = 0;
+    for (let i = view.length - 1; i >= 0; i--) {
+      acc += defaultCountTokens(view[i].text ?? "");
+      if (acc > budget) {
+        start = Math.min(i + 1, view.length - 1);
+        break;
+      }
+    }
+    if (start > 0) view = view.slice(start);
+  }
+  return view;
+}
 function epermAlertThreshold() {
   const env = process.env.BILI_PERSIST_EPERM_ALERT_THRESHOLD;
   if (env) {
@@ -50322,7 +50504,10 @@ function peekSession(id) {
   return sessions.get(id);
 }
 function snapshotMessages(session, messages) {
-  if (messages.length > 0) session.lastMessages = messages;
+  if (messages.length > 0) {
+    session.lastMessages = messages;
+    session.lastMessagesFolded = false;
+  }
 }
 function markDirty(session) {
   getStore().scheduleSave(session);
@@ -50335,6 +50520,8 @@ function resetSessionCompression(session) {
   session.blockContents.clear();
   session.stats.lastInputTokens = 0;
   session.stats.contextTokens = 0;
+  session.hostCreditTokens = 0;
+  session.hostContextTokens = 0;
   session.metadata.nativeCompactionAt = Date.now();
   markDirty(session);
 }
@@ -50448,6 +50635,50 @@ function withStagedCompressGuidance(text) {
   return text + STAGED_COMPRESS_GUIDANCE;
 }
 
+// src/acp-status.ts
+function handleAcpStatus(args, ctx) {
+  const scope = typeof args.scope === "string" ? args.scope : void 0;
+  const view = typeof args.view === "string" ? args.view : void 0;
+  const tool = typeof args.tool === "string" ? args.tool : void 0;
+  const sort = typeof args.sort === "string" ? args.sort : void 0;
+  const limit = typeof args.limit === "number" ? args.limit : void 0;
+  const base = buildStatusReport(ctx.session.state, ctx.messages, defaultCountTokens, { scope, view, tool, sort, limit });
+  if (scope) return base;
+  const extra = [];
+  try {
+    const turn = ctx.core.processTurn({
+      messages: ctx.messages,
+      state: ctx.session.state,
+      config: ctx.config,
+      tokenCount: ctx.session.stats.lastInputTokens,
+      renderTags: "none"
+    });
+    const nudge = turn.nudge;
+    if (nudge) {
+      extra.push("");
+      extra.push(nudge.shouldInject ? `Nudge: ACTIVE \u2014 ${nudge.reason}` : `Nudge: idle \u2014 ${nudge.reason}`);
+      const ranges = viableRanges(nudge.compressibleRanges);
+      const protectedRanges = nudge.protectedRanges ?? [];
+      if (ranges.length > 0 || protectedRanges.length > 0) {
+        extra.push("");
+        extra.push(formatRanges(ranges, protectedRanges));
+      }
+    }
+  } catch {
+  }
+  const archive = preCompactionArchiveOf(ctx.session);
+  const archivedIds = Object.keys(archive);
+  if (archivedIds.length > 0) {
+    extra.push("");
+    extra.push(`PRE-COMPACTION ARCHIVE \u2014 ${archivedIds.length} block(s): content was replaced by the client's native compaction summary, so it is no longer in the session history and decompress is unavailable.`);
+    for (const id of archivedIds) {
+      extra.push(`  ${id} \u2014 ${archive[id].reason}`);
+    }
+  }
+  return extra.length > 0 ? `${base}
+${extra.join("\n")}` : base;
+}
+
 // src/decompress-shared.ts
 import { mkdirSync as mkdirSync4, unlinkSync as unlinkSync2, writeFileSync as writeFileSync3 } from "fs";
 import { dirname as dirname2, join as join2 } from "path";
@@ -50769,7 +51000,7 @@ function executeAnthropicProxyTool(toolName, args, ctx) {
 ${lines.join("\n\n")}`;
   }
   if (toolName === "acp_status") {
-    return buildStatusReport(ctx.session.state, ctx.messages, defaultCountTokens);
+    return handleAcpStatus(args, ctx);
   }
   return `[Unknown proxy tool: ${toolName}]`;
 }
@@ -50876,6 +51107,7 @@ var CHUNK_FRACTION = 0.6;
 var MIN_CHUNK_TOKENS = 2e3;
 var MIN_SUMMARY_CHARS = 50;
 var MAX_SUMMARY_OUTPUT_TOKENS = 8192;
+var MAX_SUMMARY_CALLS_PER_PREFLIGHT = 8;
 function refMaps(messages, state) {
   const refToIdx = /* @__PURE__ */ new Map();
   const idxToRef = /* @__PURE__ */ new Map();
@@ -50896,6 +51128,25 @@ function estimateCoreMessages(messages) {
   for (const m2 of messages) tokens += defaultCountTokens(m2.text ?? "");
   return tokens;
 }
+var NON_TEXT_BODY_KEYS = /* @__PURE__ */ new Set(["data", "url", "b64_json", "file_data"]);
+function estimateRawBodyTokens(parsed) {
+  let tokens = 0;
+  const walk = (value, key) => {
+    if (typeof value === "string") {
+      if (!key || !NON_TEXT_BODY_KEYS.has(key)) tokens += defaultCountTokens(value);
+      return;
+    }
+    if (Array.isArray(value)) {
+      for (const item of value) walk(item, key);
+      return;
+    }
+    if (value && typeof value === "object") {
+      for (const [k2, v2] of Object.entries(value)) walk(v2, k2);
+    }
+  };
+  walk(parsed);
+  return tokens;
+}
 function rangeChars2(messages, startIdx, endIdx) {
   let chars = 0;
   for (let i = startIdx; i <= endIdx && i < messages.length; i++) {
@@ -51003,15 +51254,28 @@ TASK: The conversation segment below (messages ${startRef}\u2013${endRef}) must
   }
 }
 var ABORTED_FAILURE = { kind: "aborted", detail: "the client disconnected during preflight compression" };
+function relaxedConfig(config) {
+  return { ...config, preserveRecentMessages: 0, preserveRecentTokens: 0 };
+}
+function noEmergencyTruncate(config) {
+  return { ...config, modelContextLimit: config.modelContextLimit * 100 };
+}
 async function preflightCompress(deps, messages) {
   const limit = deps.config.modelContextLimit;
-  const result = { compressedRanges: 0, savedTokens: 0, payloadEstimate: estimateCoreMessages(messages) + (deps.imageFloor ?? 0) };
+  const result = { compressedRanges: 0, savedTokens: 0, payloadEstimate: estimateCoreMessages(messages) + (deps.imageFloor ?? 0) + (deps.wireOverhead ?? 0) };
   if (limit <= 0) return result;
   const budget = Math.max(MIN_CHUNK_TOKENS, Math.floor(limit * CHUNK_FRACTION));
   const minChars = deps.config.compress.minCompressRange;
   let currentTokens = deps.session.stats.lastInputTokens;
   let startTokens = -1;
   let failure;
+  let activeConfig = deps.config;
+  let relaxed = false;
+  const relaxedExhaustedDetail = `the payload still exceeds the window after folding everything compressible, including the soft-protected recent zone (last ${deps.config.preserveRecentMessages} messages + most recent user message), which was relaxed under overflow; hard protectedTools remain excluded. Raise the model context window or restart the session to recover.`;
+  const skipSet = /* @__PURE__ */ new Set();
+  let summaryCalls = 0;
+  let budgetHit = false;
+  let rangesTried = 0;
   for (let round = 0; round < MAX_PREFLIGHT_ROUNDS; round++) {
     if (deps.signal?.aborted) {
       failure = ABORTED_FAILURE;
@@ -51020,91 +51284,127 @@ async function preflightCompress(deps, messages) {
     const turn = deps.core.processTurn({
       messages,
       state: deps.session.state,
-      config: deps.config,
+      config: noEmergencyTruncate(activeConfig),
       tokenCount: currentTokens,
       renderTags: "text-only"
     });
     deps.session.state = turn.state;
-    currentTokens = Math.max(deps.session.stats.lastInputTokens, estimateCoreMessages(turn.messages) + (deps.imageFloor ?? 0));
-    result.payloadEstimate = estimateCoreMessages(turn.messages) + (deps.imageFloor ?? 0);
+    currentTokens = Math.max(deps.session.stats.lastInputTokens, estimateCoreMessages(turn.messages) + (deps.imageFloor ?? 0) + (deps.wireOverhead ?? 0));
+    result.payloadEstimate = estimateCoreMessages(turn.messages) + (deps.imageFloor ?? 0) + (deps.wireOverhead ?? 0);
     if (startTokens < 0) startTokens = currentTokens;
     if (currentTokens < limit) break;
     const ranges = viableRanges(turn.nudge?.compressibleRanges ?? []);
     if (ranges.length === 0) {
-      failure = { kind: "exhausted", detail: "no compressible ranges remain in the conversation" };
-      break;
-    }
-    const range = [...ranges].sort((a, b2) => refNum3(a.startRef) - refNum3(b2.startRef))[0];
-    const { refToIdx } = refMaps(messages, deps.session.state);
-    const startIdx = refToIdx.get(range.startRef);
-    const endIdx = refToIdx.get(range.endRef);
-    if (startIdx === void 0 || endIdx === void 0 || startIdx > endIdx) {
-      failure = { kind: "exhausted", detail: "the compressible range no longer resolves to payload messages" };
+      if (!relaxed && result.payloadEstimate >= limit) {
+        activeConfig = relaxedConfig(deps.config);
+        relaxed = true;
+        summaryCalls = 0;
+        budgetHit = false;
+        deps.log("warn", "[preflight] no compressible ranges outside the protected recent zone; relaxing soft protection (preserveRecentMessages/Tokens -> 0) and retrying");
+        continue;
+      }
+      failure = { kind: "exhausted", detail: relaxed ? relaxedExhaustedDetail : "no compressible ranges remain in the conversation" };
       break;
     }
+    const ordered = [...ranges].sort((a, b2) => refNum3(a.startRef) - refNum3(b2.startRef));
     let appliedThisRound = 0;
-    for (const [cs2, ce2] of splitChunks(messages, startIdx, endIdx, budget)) {
+    for (const range of ordered) {
       if (currentTokens < limit) break;
       if (deps.signal?.aborted) {
         failure = ABORTED_FAILURE;
         break;
       }
-      const maps = refMaps(messages, deps.session.state);
-      const startRef = maps.idxToRef.get(cs2);
-      const endRef = maps.idxToRef.get(ce2);
-      if (!startRef || !endRef) continue;
-      if (rangeChars2(messages, cs2, ce2) < minChars) continue;
-      const content = renderRange(messages, cs2, ce2);
-      if (content.length === 0) continue;
-      let summary;
-      try {
-        summary = await summarizeRange(deps, content, startRef, endRef);
-      } catch (err2) {
-        if (err2 instanceof UpstreamHttpError) {
-          failure = {
-            kind: "upstream",
-            status: err2.status,
-            detail: err2.status === 429 ? `the summarization call was rate-limited by the upstream (HTTP 429)` : `the summarization call was rejected by the upstream (HTTP ${err2.status})`
-          };
-          deps.log("warn", `[preflight] summarization failed: HTTP ${err2.status} ${err2.body.slice(0, 200)}`);
-        } else if (deps.signal?.aborted) {
+      if (budgetHit) break;
+      const skipKey = `${range.startRef}:${range.endRef}`;
+      if (skipSet.has(skipKey)) continue;
+      const { refToIdx } = refMaps(messages, deps.session.state);
+      const startIdx = refToIdx.get(range.startRef);
+      const endIdx = refToIdx.get(range.endRef);
+      if (startIdx === void 0 || endIdx === void 0 || startIdx > endIdx) {
+        skipSet.add(skipKey);
+        continue;
+      }
+      rangesTried += 1;
+      for (const [cs2, ce2] of splitChunks(messages, startIdx, endIdx, budget)) {
+        if (currentTokens < limit) break;
+        if (deps.signal?.aborted) {
           failure = ABORTED_FAILURE;
-          deps.log("warn", `[preflight] summarization aborted: client disconnected`);
-        } else {
-          failure = { kind: "upstream", detail: `the summarization call failed: ${String(err2)}` };
-          deps.log("warn", `[preflight] summarization failed: ${String(err2)}`);
+          break;
         }
+        if (budgetHit) break;
+        const maps = refMaps(messages, deps.session.state);
+        const startRef = maps.idxToRef.get(cs2);
+        const endRef = maps.idxToRef.get(ce2);
+        if (!startRef || !endRef) continue;
+        if (rangeChars2(messages, cs2, ce2) < minChars) continue;
+        const content = renderRange(messages, cs2, ce2);
+        if (content.length === 0) continue;
+        if (summaryCalls >= MAX_SUMMARY_CALLS_PER_PREFLIGHT) {
+          budgetHit = true;
+          break;
+        }
+        summaryCalls += 1;
+        let summary;
+        try {
+          summary = await summarizeRange(deps, content, startRef, endRef);
+        } catch (err2) {
+          if (err2 instanceof UpstreamHttpError) {
+            failure = {
+              kind: "upstream",
+              status: err2.status,
+              detail: err2.status === 429 ? `the summarization call was rate-limited by the upstream (HTTP 429)` : `the summarization call was rejected by the upstream (HTTP ${err2.status})`
+            };
+            deps.log("warn", `[preflight] summarization failed: HTTP ${err2.status} ${err2.body.slice(0, 200)}`);
+          } else if (deps.signal?.aborted) {
+            failure = ABORTED_FAILURE;
+            deps.log("warn", `[preflight] summarization aborted: client disconnected`);
+          } else {
+            failure = { kind: "upstream", detail: `the summarization call failed: ${String(err2)}` };
+            deps.log("warn", `[preflight] summarization failed: ${String(err2)}`);
+          }
+          break;
+        }
+        if (!summary) {
+          deps.log("warn", `[preflight] range ${skipKey} produced no usable summary; skipping it`);
+          skipSet.add(skipKey);
+          break;
+        }
+        const ctx = {
+          core: deps.core,
+          config: activeConfig,
+          messages,
+          session: deps.session,
+          log: (msg) => deps.log("info", msg)
+        };
+        const creditBefore = deps.session.stats.compressCreditTokens;
+        const applied = applyRanges(parseCompressInput({ content: [{ startId: startRef, endId: endRef, summary, topic: "preflight overflow compress" }] }), ctx);
+        if (applied.startsWith("[Compression FAILED")) {
+          deps.log("warn", `[preflight] ${applied}`);
+          skipSet.add(skipKey);
+          break;
+        }
+        const compressed = deps.session.stats.compressCreditTokens - creditBefore;
+        currentTokens = Math.max(0, currentTokens - compressed + defaultCountTokens(summary));
+        deps.session.stats.lastInputTokens += defaultCountTokens(summary);
+        appliedThisRound += 1;
+        result.compressedRanges += 1;
         break;
       }
-      if (!summary) continue;
-      const ctx = {
-        core: deps.core,
-        config: deps.config,
-        messages,
-        session: deps.session,
-        log: (msg) => deps.log("info", msg)
-      };
-      const creditBefore = deps.session.stats.compressCreditTokens;
-      const applied = applyRanges(parseCompressInput({ content: [{ startId: startRef, endId: endRef, summary, topic: "preflight overflow compress" }] }), ctx);
-      if (applied.startsWith("[Compression FAILED")) {
-        deps.log("warn", `[preflight] ${applied}`);
-        continue;
-      }
-      const compressed = deps.session.stats.compressCreditTokens - creditBefore;
-      currentTokens = Math.max(0, currentTokens - compressed + defaultCountTokens(summary));
-      deps.session.stats.lastInputTokens += defaultCountTokens(summary);
-      appliedThisRound += 1;
-      result.compressedRanges += 1;
-    }
-    if (appliedThisRound === 0) {
-      if (!failure) {
-        failure = { kind: "exhausted", detail: "no range could be compressed (chunks below minCompressRange or the summarization responses were unusable)" };
-      }
-      break;
+      if (appliedThisRound > 0) break;
+      if (failure || budgetHit) break;
     }
+    if (appliedThisRound === 0) break;
   }
-  if (!failure && currentTokens >= limit) {
-    failure = { kind: "exhausted", detail: `the compress budget was exhausted after ${MAX_PREFLIGHT_ROUNDS} rounds` };
+  if (currentTokens >= limit && !failure) {
+    if (budgetHit) {
+      failure = { kind: "exhausted", detail: `the preflight summarization budget (${MAX_SUMMARY_CALLS_PER_PREFLIGHT} calls per protection regime) was exhausted before the payload fit the window` };
+    } else if (relaxed && result.compressedRanges > 0) {
+      failure = { kind: "exhausted", detail: relaxedExhaustedDetail };
+    } else if (result.compressedRanges === 0) {
+      failure = { kind: "exhausted", detail: `no range could be compressed across ${rangesTried} viable range${rangesTried === 1 ? "" : "s"} (each was below minCompressRange, had an unusable summary, or failed to apply)` };
+    } else {
+      failure = { kind: "exhausted", detail: `the compress budget was exhausted after ${MAX_PREFLIGHT_ROUNDS} rounds` };
+    }
   }
   if (result.compressedRanges > 0) deps.session.stats.lastInputTokens = currentTokens;
   result.savedTokens = Math.max(0, startTokens - currentTokens);
@@ -51634,7 +51934,7 @@ function reapOrphanBlocks(session, visible, deactivate) {
 }
 
 // src/instance.ts
-import { randomUUID as randomUUID3 } from "crypto";
+import { createHash as createHash4, randomUUID as randomUUID3 } from "crypto";
 import fs3 from "fs";
 import path7 from "path";
 function isProxyInstanceFile(v2) {
@@ -51680,14 +51980,13 @@ function readProxyInstanceFile(file) {
 function instanceFilePath() {
   return path7.join(stateDir(), "proxy-origin");
 }
-function atomicWriteInstanceFile(info, file) {
-  const filePath = file ?? instanceFilePath();
+function atomicWriteJson(obj, filePath) {
   fs3.mkdirSync(path7.dirname(filePath), { recursive: true });
   const tempPath = `${filePath}.${process.pid}.${randomUUID3()}.tmp`;
   let descriptor;
   try {
     descriptor = fs3.openSync(tempPath, "wx", 420);
-    fs3.writeSync(descriptor, JSON.stringify(info) + "\n", null, "utf8");
+    fs3.writeSync(descriptor, JSON.stringify(obj) + "\n", null, "utf8");
     fs3.fsyncSync(descriptor);
     fs3.closeSync(descriptor);
     descriptor = void 0;
@@ -51706,6 +52005,9 @@ function atomicWriteInstanceFile(info, file) {
     throw error;
   }
 }
+function atomicWriteInstanceFile(info, file) {
+  atomicWriteJson(info, file ?? instanceFilePath());
+}
 function clearProxyInstanceFile(instanceId, file) {
   const filePath = file ?? instanceFilePath();
   const current = readProxyInstanceFile(filePath);
@@ -51725,44 +52027,99 @@ function isPidAlive(pid) {
     return err2.code === "EPERM";
   }
 }
-function registryFilePath() {
+function registryDirPath() {
+  return path7.join(stateDir(), "instances");
+}
+function legacyRegistryFilePath() {
   return path7.join(stateDir(), "instances.json");
 }
-function registerInstanceAndWarn(entry, warn) {
-  const file = registryFilePath();
-  const entries = [];
+var SAFE_NAME_RE = /^[A-Za-z0-9][A-Za-z0-9._-]*$/;
+function safeRegistryName(instanceId) {
+  if (instanceId !== "." && instanceId !== ".." && instanceId.length <= 128 && SAFE_NAME_RE.test(instanceId)) {
+    return instanceId;
+  }
+  return createHash4("sha256").update(instanceId).digest("hex");
+}
+function registryEntryFile(instanceId) {
+  return path7.join(registryDirPath(), `${safeRegistryName(instanceId)}.json`);
+}
+function safeReadJson2(file) {
   try {
-    const parsed = JSON.parse(fs3.readFileSync(file, "utf8"));
-    if (Array.isArray(parsed.instances)) {
-      for (const e of parsed.instances) {
-        if (e && typeof e.instanceId === "string" && e.instanceId !== entry.instanceId && isPidAlive(e.pid)) {
-          entries.push(e);
-        }
+    return JSON.parse(fs3.readFileSync(file, "utf8"));
+  } catch {
+    return void 0;
+  }
+}
+function coerceEntry(value) {
+  if (!value || typeof value !== "object") return void 0;
+  const o = value;
+  if (typeof o.instanceId !== "string" || o.instanceId === "") return void 0;
+  return {
+    instanceId: o.instanceId,
+    pid: typeof o.pid === "number" ? o.pid : 0,
+    port: typeof o.port === "number" ? o.port : 0,
+    origin: typeof o.origin === "string" ? o.origin : "",
+    startedAt: typeof o.startedAt === "number" ? o.startedAt : 0
+  };
+}
+function readMarkerNames() {
+  try {
+    return fs3.readdirSync(registryDirPath());
+  } catch {
+    return [];
+  }
+}
+function readAllRegistryEntries() {
+  const seen = /* @__PURE__ */ new Set();
+  const out = [];
+  for (const name of readMarkerNames()) {
+    if (!name.endsWith(".json")) continue;
+    const entry = coerceEntry(safeReadJson2(path7.join(registryDirPath(), name)));
+    if (entry && !seen.has(entry.instanceId)) {
+      seen.add(entry.instanceId);
+      out.push(entry);
+    }
+  }
+  const legacy = safeReadJson2(legacyRegistryFilePath());
+  if (legacy && Array.isArray(legacy.instances)) {
+    for (const raw of legacy.instances) {
+      const entry = coerceEntry(raw);
+      if (entry && !seen.has(entry.instanceId)) {
+        seen.add(entry.instanceId);
+        out.push(entry);
       }
     }
-  } catch {
   }
-  for (const other of entries) {
+  return out;
+}
+function reapDeadMarkers(ours) {
+  for (const name of readMarkerNames()) {
+    if (!name.endsWith(".json")) continue;
+    const file = path7.join(registryDirPath(), name);
+    const entry = coerceEntry(safeReadJson2(file));
+    if (!entry || entry.instanceId === ours || isPidAlive(entry.pid)) continue;
+    try {
+      fs3.unlinkSync(file);
+    } catch {
+    }
+  }
+}
+function registerInstanceAndWarn(entry, warn) {
+  const others = readAllRegistryEntries().filter((e) => e.instanceId !== entry.instanceId && isPidAlive(e.pid));
+  for (const other of others) {
     warn(
       `another bili instance is running (pid ${other.pid}, ${other.origin}) \u2014 both processes will write the same sessions directory; stop one to avoid state pollution (#394)`
     );
   }
-  entries.push(entry);
+  reapDeadMarkers(entry.instanceId);
   try {
-    fs3.mkdirSync(path7.dirname(file), { recursive: true });
-    fs3.writeFileSync(file, JSON.stringify({ instances: entries }) + "\n");
+    atomicWriteJson(entry, registryEntryFile(entry.instanceId));
   } catch {
   }
 }
 function unregisterInstance(instanceId) {
-  const file = registryFilePath();
   try {
-    const parsed = JSON.parse(fs3.readFileSync(file, "utf8"));
-    if (!Array.isArray(parsed.instances)) return;
-    const kept = parsed.instances.filter(
-      (e) => !(e && typeof e.instanceId === "string" && e.instanceId === instanceId) && isPidAlive(e.pid)
-    );
-    fs3.writeFileSync(file, JSON.stringify({ instances: kept }) + "\n");
+    fs3.unlinkSync(registryEntryFile(instanceId));
   } catch {
   }
 }
@@ -51879,6 +52236,126 @@ function warnCacheCollapse(session, input, cached) {
   );
 }
 
+// src/util.ts
+import { createHash as createHash5 } from "crypto";
+function hashId2(s3) {
+  return createHash5("sha256").update(s3, "utf8").digest("hex").slice(0, 16);
+}
+function isLoopbackAddress(addr) {
+  return !!addr && (addr.startsWith("127.") || addr === "::1" || addr.startsWith("::ffff:127."));
+}
+function usageTotals(protocol, usage) {
+  const num3 = (v2) => typeof v2 === "number" && Number.isFinite(v2) ? v2 : void 0;
+  if (protocol === "anthropic") {
+    const fresh = num3(usage["input_tokens"]);
+    const read = num3(usage["cache_read_input_tokens"]);
+    const creation = num3(usage["cache_creation_input_tokens"]);
+    const any = fresh !== void 0 || read !== void 0 || creation !== void 0;
+    return {
+      total: any ? (fresh ?? 0) + (read ?? 0) + (creation ?? 0) : void 0,
+      cached: read
+    };
+  }
+  if (protocol === "openai") {
+    const prompt = num3(usage["prompt_tokens"]);
+    const cached = num3(usage["prompt_tokens_details"]?.["cached_tokens"]);
+    return {
+      total: prompt !== void 0 ? promptInputTotal("openai", prompt, cached) : void 0,
+      cached
+    };
+  }
+  return {
+    total: num3(usage["input_tokens"]),
+    cached: num3(usage["input_tokens_details"]?.["cached_tokens"])
+  };
+}
+function promptInputTotal(protocol, input, cached) {
+  if (input === void 0) return 0;
+  const includesCached = protocol === "openai" || protocol === "responses";
+  const splitSemantics = !includesCached || typeof cached === "number" && input < cached;
+  return input + (splitSemantics && typeof cached === "number" ? cached : 0);
+}
+function backfillHostUsage(protocol, usage, credit) {
+  if (!Number.isFinite(credit) || credit <= 0) return false;
+  let patched = false;
+  const add = (key) => {
+    if (typeof usage[key] === "number" && Number.isFinite(usage[key])) {
+      usage[key] = usage[key] + credit;
+      patched = true;
+    }
+  };
+  if (protocol === "openai") {
+    add("prompt_tokens");
+    add("total_tokens");
+  } else {
+    add("input_tokens");
+  }
+  return patched;
+}
+var CONTEXT_OVERFLOW_PATTERNS = [
+  /context_length_exceeded/i,
+  /context_window_exceeded/i,
+  /context length exceeded/i,
+  /maximum context length/i,
+  /max context length/i,
+  /maximum context size/i,
+  /longer than the model'?s context length/i,
+  /exceeds the context window/i,
+  /out of room in the model/i,
+  /exceeded model token limit/i,
+  /prompt is too long/i,
+  /prompt_too_long/i,
+  /prompt_is_too_long/i,
+  /request_too_large/i,
+  /token limit exceeded/i,
+  // #554: llama.cpp-family "exceed_context_size_error (A / B > W)" — carried by
+  // side requests that bypass preflight; without it the learned channel learns nothing.
+  /exceed[_\s]?context[_\s]?size/i
+];
+function toTokenNumber(s3) {
+  const n = parseInt(s3.replace(/,/g, ""), 10);
+  return Number.isFinite(n) && n >= 1e3 ? n : void 0;
+}
+function parseOverflowWindow(text) {
+  let m2 = text.match(/>\s*(\d[\d,]*)\s*maximum/i);
+  if (m2) return toTokenNumber(m2[1]);
+  m2 = text.match(/\(\s*\d[\d,]*\s*\/\s*\d[\d,]*\s*>\s*(\d[\d,]+)\s*\)/);
+  if (m2) return toTokenNumber(m2[1]);
+  m2 = text.match(/maximum context length is (\d[\d,]*)/i) ?? text.match(/maximum context length of (\d[\d,]*)/i) ?? text.match(/maximum context size (?:is|of) (\d[\d,]*)/i) ?? text.match(/(?:maximum|max)\s+(?:context\s+)?length\s+(?:is\s+)?(\d[\d,]*)/i) ?? text.match(/context length\s*\((\d[\d,]*)\s*token/i) ?? text.match(/limit of (\d[\d,]*)\s*token/i) ?? text.match(/(\d[\d,]*)\s*maximum\b/i);
+  if (m2) return toTokenNumber(m2[1]);
+  return void 0;
+}
+function inspectContextOverflow(status, bodyText) {
+  const message = (bodyText ?? "").slice(0, 300);
+  if (status !== 400 && status !== 413) return { isOverflow: false, message };
+  if (!bodyText) return { isOverflow: false, message };
+  const isOverflow = CONTEXT_OVERFLOW_PATTERNS.some((p2) => p2.test(bodyText));
+  if (!isOverflow) return { isOverflow: false, message };
+  return { isOverflow: true, window: parseOverflowWindow(bodyText), message };
+}
+function reserveOutputHeadroom(window2, maxOutput) {
+  if (Number.isFinite(window2) && window2 > 0 && Number.isFinite(maxOutput) && maxOutput > 0 && maxOutput < window2) {
+    return window2 - maxOutput;
+  }
+  return window2;
+}
+function systemToUser(messages) {
+  let hasSys = false;
+  for (const m2 of messages) {
+    if (m2.role === "system" || m2.role === "developer") {
+      hasSys = true;
+      break;
+    }
+  }
+  if (!hasSys) return messages;
+  return messages.map(
+    (m2) => m2.role === "system" || m2.role === "developer" ? { ...m2, role: "user" } : m2
+  );
+}
+function shouldReserveOutputHeadroom(protocol) {
+  return protocol !== "anthropic";
+}
+
 // src/loop/core.ts
 var MAX_LOOP_ROUNDS = 10;
 function isLoopThinking(m2) {
@@ -51932,46 +52409,14 @@ ${lines.join("\n\n")}`;
   }
   return `[Unknown proxy tool: ${toolName}]`;
 }
-function handleAcpStatus(args, ctx) {
-  const scope = typeof args.scope === "string" ? args.scope : void 0;
-  const view = typeof args.view === "string" ? args.view : void 0;
-  const tool = typeof args.tool === "string" ? args.tool : void 0;
-  const sort = typeof args.sort === "string" ? args.sort : void 0;
-  const limit = typeof args.limit === "number" ? args.limit : void 0;
-  const base = buildStatusReport(ctx.session.state, ctx.messages, defaultCountTokens, { scope, view, tool, sort, limit });
-  if (scope) return base;
-  const nudge = ctx.nudge;
-  const ranges = nudge?.compressibleRanges ?? [];
-  const protectedRanges = nudge?.protectedRanges ?? [];
-  const archive = preCompactionArchiveOf(ctx.session);
-  const archivedIds = Object.keys(archive);
-  const extra = [];
-  if (nudge) {
-    extra.push("");
-    extra.push(nudge.shouldInject ? `Nudge: ACTIVE \u2014 ${nudge.reason}` : `Nudge: idle \u2014 ${nudge.reason}`);
-  }
-  if (ranges.length > 0 || protectedRanges.length > 0) {
-    extra.push("");
-    extra.push(formatRanges(ranges, protectedRanges));
-  }
-  if (archivedIds.length > 0) {
-    extra.push("");
-    extra.push(`PRE-COMPACTION ARCHIVE \u2014 ${archivedIds.length} block(s): content was replaced by the client's native compaction summary, so it is no longer in the session history and decompress is unavailable.`);
-    for (const id of archivedIds) {
-      extra.push(`  ${id} \u2014 ${archive[id].reason}`);
-    }
-  }
-  return extra.length > 0 ? `${base}
-${extra.join("\n")}` : base;
-}
 function recordUsage(ctx, usage, round) {
   const prompt = usage.inputTokens;
   const cached = usage.cachedTokens;
   const out = usage.outputTokens;
-  const includesCached = ctx.protocol === "openai" || ctx.protocol === "responses";
-  const total = (typeof prompt === "number" ? prompt : 0) + (!includesCached && typeof cached === "number" ? cached : 0);
+  const total = promptInputTotal(ctx.protocol, prompt, cached);
   if (total > 0) ctx.session.stats.inputTokens += total;
   ctx.session.stats.lastInputTokens = Math.max(0, total - (ctx.session.stats.compressCreditTokens ?? 0));
+  ctx.session.hostContextTokens = total + (ctx.session.hostCreditTokens ?? 0);
   if (typeof cached === "number") {
     ctx.session.stats.cachedTokens += cached;
     ctx.session.stats.cacheSamples += 1;
@@ -51995,7 +52440,7 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
     {
       method: "POST",
       headers: requestOptions.headers,
-      body: JSON.stringify(body),
+      body: JSON.stringify(requestOptions.wireTransform ? requestOptions.wireTransform(body) : body),
       ...ctx.proxyUrl ? { dispatcher: proxyDispatcher(ctx.proxyUrl) } : {}
     },
     void 0,
@@ -52110,6 +52555,10 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
       if (usage.inputTokens !== void 0 || usage.outputTokens !== void 0 || usage.cachedTokens !== void 0) {
         recordUsage(ctx, usage, round);
       }
+      const hostCredit = ctx.session.hostCreditTokens ?? 0;
+      if (hostCredit > 0 && typeof usage.inputTokens === "number") {
+        usage.inputTokens += hostCredit;
+      }
       let resolvedText = assistantText;
       let allCalls = calls;
       if (ctx.textProtocol && assistantText.length > 0 && adapter.extractTextTriggers) {
@@ -52337,96 +52786,6 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
   }
 }
 
-// src/util.ts
-import { createHash as createHash4 } from "crypto";
-function hashId2(s3) {
-  return createHash4("sha256").update(s3, "utf8").digest("hex").slice(0, 16);
-}
-function isLoopbackAddress(addr) {
-  return !!addr && (addr.startsWith("127.") || addr === "::1" || addr.startsWith("::ffff:127."));
-}
-function usageTotals(protocol, usage) {
-  const num3 = (v2) => typeof v2 === "number" && Number.isFinite(v2) ? v2 : void 0;
-  if (protocol === "anthropic") {
-    const fresh = num3(usage["input_tokens"]);
-    const read = num3(usage["cache_read_input_tokens"]);
-    const creation = num3(usage["cache_creation_input_tokens"]);
-    const any = fresh !== void 0 || read !== void 0 || creation !== void 0;
-    return {
-      total: any ? (fresh ?? 0) + (read ?? 0) + (creation ?? 0) : void 0,
-      cached: read
-    };
-  }
-  if (protocol === "openai") {
-    return {
-      total: num3(usage["prompt_tokens"]),
-      cached: num3(usage["prompt_tokens_details"]?.["cached_tokens"])
-    };
-  }
-  return {
-    total: num3(usage["input_tokens"]),
-    cached: num3(usage["input_tokens_details"]?.["cached_tokens"])
-  };
-}
-var CONTEXT_OVERFLOW_PATTERNS = [
-  /context_length_exceeded/i,
-  /context_window_exceeded/i,
-  /context length exceeded/i,
-  /maximum context length/i,
-  /max context length/i,
-  /maximum context size/i,
-  /longer than the model'?s context length/i,
-  /exceeds the context window/i,
-  /out of room in the model/i,
-  /exceeded model token limit/i,
-  /prompt is too long/i,
-  /prompt_too_long/i,
-  /prompt_is_too_long/i,
-  /request_too_large/i,
-  /token limit exceeded/i
-];
-function toTokenNumber(s3) {
-  const n = parseInt(s3.replace(/,/g, ""), 10);
-  return Number.isFinite(n) && n >= 1e3 ? n : void 0;
-}
-function parseOverflowWindow(text) {
-  let m2 = text.match(/>\s*(\d[\d,]*)\s*maximum/i);
-  if (m2) return toTokenNumber(m2[1]);
-  m2 = text.match(/maximum context length is (\d[\d,]*)/i) ?? text.match(/maximum context length of (\d[\d,]*)/i) ?? text.match(/maximum context size (?:is|of) (\d[\d,]*)/i) ?? text.match(/(?:maximum|max)\s+(?:context\s+)?length\s+(?:is\s+)?(\d[\d,]*)/i) ?? text.match(/context length\s*\((\d[\d,]*)\s*token/i) ?? text.match(/limit of (\d[\d,]*)\s*token/i) ?? text.match(/(\d[\d,]*)\s*maximum\b/i);
-  if (m2) return toTokenNumber(m2[1]);
-  return void 0;
-}
-function inspectContextOverflow(status, bodyText) {
-  const message = (bodyText ?? "").slice(0, 300);
-  if (status !== 400 && status !== 413) return { isOverflow: false, message };
-  if (!bodyText) return { isOverflow: false, message };
-  const isOverflow = CONTEXT_OVERFLOW_PATTERNS.some((p2) => p2.test(bodyText));
-  if (!isOverflow) return { isOverflow: false, message };
-  return { isOverflow: true, window: parseOverflowWindow(bodyText), message };
-}
-function reserveOutputHeadroom(window2, maxOutput) {
-  if (Number.isFinite(window2) && window2 > 0 && Number.isFinite(maxOutput) && maxOutput > 0 && maxOutput < window2) {
-    return window2 - maxOutput;
-  }
-  return window2;
-}
-function systemToUser(messages) {
-  let hasSys = false;
-  for (const m2 of messages) {
-    if (m2.role === "system" || m2.role === "developer") {
-      hasSys = true;
-      break;
-    }
-  }
-  if (!hasSys) return messages;
-  return messages.map(
-    (m2) => m2.role === "system" || m2.role === "developer" ? { ...m2, role: "user" } : m2
-  );
-}
-function shouldReserveOutputHeadroom(protocol) {
-  return protocol !== "anthropic";
-}
-
 // src/loop/adapter-responses.ts
 var RESPONSES_ITEM_ID_MAX = 64;
 function normalizeResponsesMessageItems(input) {
@@ -53058,7 +53417,7 @@ function stripFinishReasonChunk(buf) {
     return buf;
   }
 }
-function createOpenaiAdapter(requestBody, clientSystem) {
+function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0) {
   const model = requestBody.model ?? "unknown";
   let responseId = `chatcmpl-proxy-${Date.now()}`;
   let toolIndex = 0;
@@ -53252,7 +53611,24 @@ function createOpenaiAdapter(requestBody, clientSystem) {
             cachedTokens: typeof pd?.cached_tokens === "number" ? pd.cached_tokens : void 0
           };
           if (sawRealToolCall) {
-            yield { kind: "meta", chunk: rawBuf };
+            let chunk = rawBuf;
+            if (hostCredit > 0 && u2) {
+              const pu = typeof u2.prompt_tokens === "number" ? u2.prompt_tokens : void 0;
+              const tu = typeof u2.total_tokens === "number" ? u2.total_tokens : void 0;
+              if (pu !== void 0 || tu !== void 0) {
+                const patched = {
+                  ...parsed,
+                  usage: {
+                    ...u2,
+                    ...pu !== void 0 ? { prompt_tokens: pu + hostCredit } : {},
+                    ...tu !== void 0 ? { total_tokens: tu + hostCredit } : {}
+                  }
+                };
+                const out = eventStr.split("\n").map((l) => l.startsWith("data:") ? `data: ${JSON.stringify(patched)}` : l).join("\n");
+                chunk = Buffer.from(out + "\n\n", "utf8");
+              }
+            }
+            yield { kind: "meta", chunk };
             yield { kind: "done", finishReason, suppressCompletion: true };
             continue;
           } else {
@@ -53439,7 +53815,7 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: "tex
     "utf8"
   );
 }
-function createAnthropicAdapter(requestBody, originalSystem) {
+function createAnthropicAdapter(requestBody, originalSystem, hostCredit = 0) {
   const model = requestBody.model ?? void 0;
   let messageId;
   let clientIndex = 0;
@@ -53553,8 +53929,19 @@ ${systemPrompt}` : systemPrompt;
           if (typeof u2.input_tokens === "number") roundInput = u2.input_tokens;
           if (typeof u2.cache_read_input_tokens === "number") roundCached = u2.cache_read_input_tokens;
           if (round === 1) {
+            let chunk = rawBuf;
+            if (hostCredit > 0 && typeof u2.input_tokens === "number") {
+              const patched = structuredClone(data);
+              const pmsg = patched["message"];
+              const pu = pmsg?.["usage"] ?? {};
+              if (typeof pu.input_tokens === "number") {
+                pu.input_tokens += hostCredit;
+                const out = eventStr.split("\n").map((l) => l.startsWith("data:") ? `data: ${JSON.stringify(patched)}` : l).join("\n");
+                chunk = Buffer.from(out + "\n\n", "utf8");
+              }
+            }
             messageStartForwarded = true;
-            yield { kind: "meta", chunk: rawBuf, firstRoundOnly: true };
+            yield { kind: "meta", chunk, firstRoundOnly: true };
           }
         } else if (type === "ping") {
           yield { kind: "meta", chunk: rawBuf };
@@ -53737,10 +54124,10 @@ data: ${JSON.stringify({ type: "content_block_stop", index })}
 }
 
 // src/loop/index.ts
-function pickAdapter(protocol, requestBody, textProtocol, responsesProjection, anthropicSystem, openaiSystem) {
+function pickAdapter(protocol, requestBody, textProtocol, responsesProjection, anthropicSystem, openaiSystem, hostCredit = 0) {
   if (protocol === "responses") return createResponsesAdapter(textProtocol, responsesProjection);
-  if (protocol === "openai") return createOpenaiAdapter(requestBody, openaiSystem);
-  if (protocol === "anthropic") return createAnthropicAdapter(requestBody, anthropicSystem);
+  if (protocol === "openai") return createOpenaiAdapter(requestBody, openaiSystem, hostCredit);
+  if (protocol === "anthropic") return createAnthropicAdapter(requestBody, anthropicSystem, hostCredit);
   throw new Error(`[acp-loop] unknown protocol: ${protocol}`);
 }
 
@@ -53801,7 +54188,7 @@ function executeProxyTool2(toolName, args, ctx) {
 ${lines.join("\n\n")}`;
   }
   if (toolName === "acp_status") {
-    return buildStatusReport(ctx.session.state, ctx.messages, defaultCountTokens);
+    return handleAcpStatus(args, ctx);
   }
   return `[Unknown proxy tool: ${toolName}]`;
 }
@@ -53898,7 +54285,7 @@ async function compressLoopResponsesJson(initialResponse, ctx, requestBody, requ
     const result = await fetchWithRetry(requestOptions.url, {
       method: "POST",
       headers: requestOptions.headers,
-      body: JSON.stringify(requestBody),
+      body: JSON.stringify(requestOptions.wireTransform ? requestOptions.wireTransform(requestBody) : requestBody),
       ...ctx.proxyUrl ? { dispatcher: proxyDispatcher(ctx.proxyUrl) } : {}
     }, void 0, void 0, (info) => {
       const lc = lastCompressSuffix(ctx.session.lastCompress);
@@ -54316,6 +54703,35 @@ data: ${JSON.stringify({ type: "message_delta", delta: { stop_reason: "end_turn"
       safeWrite(res, `event: message_stop
 data: ${JSON.stringify({ type: "message_stop" })}
 
+`);
+    }
+  } catch {
+  } finally {
+    try {
+      res.end();
+    } catch {
+    }
+  }
+}
+function emitPreflightError(res, protocol, error, log2) {
+  const err2 = { type: "server_error", code: "preflight_compress_failed", message: error.message, retryable: error.retryable };
+  log2?.(`[acp-proxy: preflight failed after early response commit \u2014 delivering in-band: ${error.message}]`);
+  try {
+    if (protocol === "openai") {
+      safeWrite(res, `data: ${JSON.stringify({ error: err2 })}
+
+data: [DONE]
+
+`);
+    } else if (protocol === "responses") {
+      safeWrite(res, `event: error
+data: ${JSON.stringify({ type: "error", code: err2.code, message: err2.message })}
+
+`);
+    } else {
+      safeWrite(res, `event: error
+data: ${JSON.stringify({ type: "error", error: { type: "server_error", code: err2.code, message: err2.message } })}
+
 `);
     }
   } catch {
@@ -54371,7 +54787,7 @@ function codexTurnIdentity(headers) {
 }
 
 // src/prefix-affinity.ts
-import { createHash as createHash5 } from "crypto";
+import { createHash as createHash6 } from "crypto";
 var MIN_CANONICAL_BYTES = 24;
 var MAX_TRACKED_SESSIONS3 = 256;
 var TTL_MS2 = 7 * 24 * 60 * 60 * 1e3;
@@ -54393,7 +54809,7 @@ function sortKeys2(value) {
   return value;
 }
 function sha256(text) {
-  return createHash5("sha256").update(text, "utf8").digest("hex");
+  return createHash6("sha256").update(text, "utf8").digest("hex");
 }
 function hasUserMessage(messages) {
   return messages.some((m2) => !!m2 && typeof m2 === "object" && m2.role === "user");
@@ -55006,7 +55422,7 @@ function handlePluginManifest(res) {
     statusEndpoint: "/__bili/plugin/status"
   }));
 }
-function handlePluginStatus(conversationId2, res, fallbackLatest = false) {
+function handlePluginStatus(conversationId2, res, deps, fallbackLatest = false) {
   let entry = conversations.get(conversationId2);
   let session = entry ? peekSession(entry.sessionId) : void 0;
   let viaFallback = false;
@@ -55030,16 +55446,33 @@ function handlePluginStatus(conversationId2, res, fallbackLatest = false) {
   const limit = session.metadata.effectiveContextLimit;
   const mem = remembered.get(session.id);
   const modelContextLimit = typeof limit === "number" && limit > 0 ? limit : 0;
+  let nudge;
+  try {
+    const messages = mem ? mem.processed.length > 0 ? mem.processed : mem.original : [];
+    if (messages.length > 0) {
+      nudge = deps.core.processTurn({
+        messages,
+        state: session.state,
+        config: deps.config,
+        tokenCount: session.stats.lastInputTokens,
+        renderTags: "none"
+      }).nudge;
+    }
+  } catch {
+    nudge = void 0;
+  }
   let panel;
   try {
+    const sysTokRaw = session.metadata.systemPromptTokens;
+    const systemPromptTokens = typeof sysTokRaw === "number" && Number.isFinite(sysTokRaw) && sysTokRaw > 0 ? sysTokRaw : 0;
     panel = buildStatusPanel({
       version: `billion-context@${PROXY_VERSION}`,
-      tokenCount: session.stats.lastInputTokens,
-      systemPromptTokens: 0,
+      tokenCount: session.hostContextTokens ?? session.stats.lastInputTokens,
+      systemPromptTokens,
       state: session.state,
-      nudge: mem?.nudge,
+      nudge,
       modelContextLimit,
-      unprunedTokens: mem && mem.original.length > 0 ? mem.original.reduce((sum, m2) => sum + defaultCountTokens(m2.text ?? ""), 0) : void 0
+      unprunedTokens: mem && mem.original.length > 0 ? mem.original.reduce((sum, m2) => sum + defaultCountTokens(m2.text ?? ""), 0) + systemPromptTokens : void 0
     });
   } catch {
     panel = void 0;
@@ -55052,7 +55485,8 @@ function handlePluginStatus(conversationId2, res, fallbackLatest = false) {
     label: session.meta.label ?? null,
     pluginAgent: session.metadata.pluginAgent ?? null,
     contextLimit: typeof limit === "number" ? limit : null,
-    contextTokens: session.stats.lastInputTokens,
+    contextTokens: session.hostContextTokens ?? session.stats.lastInputTokens,
+    hostCredit: session.hostCreditTokens ?? 0,
     inputTokens: session.stats.inputTokens,
     outputTokens: session.stats.outputTokens,
     cachedTokens: session.stats.cachedTokens,
@@ -55104,8 +55538,7 @@ async function handlePluginTool(payload, res, deps) {
         config: deps.config,
         messages,
         session,
-        log: (m2) => deps.log("info", `[${session.id}] [plugin] ${m2}`),
-        nudge: mem?.nudge
+        log: (m2) => deps.log("info", `[${session.id}] [plugin] ${m2}`)
       }, callId);
     });
   } catch (err2) {
@@ -55159,16 +55592,16 @@ function usageFromSseEvent(obj) {
   return void 0;
 }
 function applyUsageSample(session, sample, protocol) {
-  const includesCached = protocol === "openai" || protocol === "responses";
   if (sample.cachedTokens !== void 0) {
     session.stats.cachedTokens += sample.cachedTokens;
     session.stats.cacheSamples += 1;
   }
   if (sample.inputTokens !== void 0) {
-    const total = sample.inputTokens + (!includesCached && sample.cachedTokens !== void 0 ? sample.cachedTokens : 0);
+    const total = promptInputTotal(protocol, sample.inputTokens, sample.cachedTokens);
     session.stats.inputTokens += total;
     session.stats.lastInputTokens = Math.max(0, total - (session.stats.compressCreditTokens ?? 0));
     warnCacheCollapse(session, total, sample.cachedTokens ?? 0);
+    session.hostContextTokens = total + (session.hostCreditTokens ?? 0);
   }
   if (sample.outputTokens !== void 0) session.stats.outputTokens += sample.outputTokens;
 }
@@ -55182,6 +55615,7 @@ async function pipePluginChatWithStrip(stream2, res, protocol, session, log2) {
   const decoder = new TextDecoder("utf-8");
   let buf = "";
   const acc = {};
+  const credit = session?.hostCreditTokens ?? 0;
   const onDrop = (snippet) => {
     log("warn", `[tag-echo] stripped model-emitted render tag (plugin passthrough): ${snippet.slice(0, 80).replace(/\n/g, " ")}`);
     log2?.(`[tag-echo] stripped model-emitted render tag from plugin passthrough text`);
@@ -55339,12 +55773,23 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
           if (ev["type"] === "message_stop") sawTerminal = true;
           const sample = usageFromSseEvent(ev);
           if (sample) mergeUsageSample(acc, sample);
+          let backfilled = false;
+          if (credit > 0 && protocol) {
+            const usage = ev["type"] === "message_start" ? ev["message"]?.["usage"] : ev["usage"];
+            const deltaEcho = ev["type"] === "message_delta" && (num2(usage?.["input_tokens"]) ?? 0) <= 0;
+            if (usage && !deltaEcho && backfillHostUsage(protocol, usage, credit)) backfilled = true;
+          }
           const out = protocol === "anthropic" ? processAnthropic(ev, rawEvent) : processOpenai(ev, rawEvent);
-          if (out.length > 0) await write(out);
+          if (backfilled && out === rawEvent + "\n\n") {
+            await write(rebuildEvent(rawEvent, ev));
+          } else if (out.length > 0) {
+            await write(out);
+          }
         }
       }
       if (res.destroyed || res.writableEnded) break;
     }
+    if (buf.length > 0 && !res.destroyed && !res.writableEnded) await write(buf);
     const rest = flushTails();
     if (rest.length > 0 && !res.destroyed && !res.writableEnded) await write(rest);
     settleUsage();
@@ -55389,6 +55834,7 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
     if (!res.write(Buffer.from(s3, "utf8"))) {
       return new Promise((r) => res.once("drain", () => r()));
     }
+    return Promise.resolve();
   };
   const settleUsage = () => {
     if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
@@ -55445,7 +55891,17 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
           const type = ev["type"];
           if (type === "response.output_text.done" || type === "response.content_part.done" || type === "response.output_item.done" || type === "response.completed" || type === "response.failed" || type === "response.incomplete") {
             if (type === "response.completed" || type === "response.failed" || type === "response.incomplete") sawTerminal = true;
-            const out = containsRenderTagText(jsonStr) ? rebuildEvent(rawEvent, stripResponsesText(ev)) : rawEvent + "\n\n";
+            let evOut = ev;
+            let rebuild = containsRenderTagText(jsonStr);
+            if (rebuild) evOut = stripResponsesText(ev);
+            if (type === "response.completed") {
+              const credit = session?.hostCreditTokens ?? 0;
+              const usage = evOut["response"]?.["usage"];
+              if (credit > 0 && usage && backfillHostUsage("responses", usage, credit)) {
+                rebuild = true;
+              }
+            }
+            const out = rebuild ? rebuildEvent(rawEvent, evOut) : rawEvent + "\n\n";
             await write(flushTail(out));
             continue;
           }
@@ -55526,8 +55982,10 @@ async function pipePluginJson(stream2, res, session, protocol) {
   }
   reader.releaseLock();
   const text = Buffer.concat(chunks).toString("utf8");
+  let json;
+  let mutated = false;
   try {
-    const json = JSON.parse(text);
+    json = JSON.parse(text);
     const usage = json["usage"];
     if (session && usage) {
       const input = num2(usage["prompt_tokens"]) ?? num2(usage["input_tokens"]);
@@ -55538,18 +55996,21 @@ async function pipePluginJson(stream2, res, session, protocol) {
           cachedTokens: num2(usage["prompt_tokens_details"]?.["cached_tokens"]) ?? num2(usage["input_tokens_details"]?.["cached_tokens"]) ?? num2(usage["cache_read_input_tokens"])
         }, protocol);
         markDirty(session);
+        const credit = session.hostCreditTokens ?? 0;
+        if (credit > 0 && protocol && backfillHostUsage(protocol, usage, credit)) {
+          mutated = true;
+        }
       }
     }
   } catch {
   }
-  if (containsRenderTagText(text)) {
-    try {
-      const json = JSON.parse(text);
-      const stripped = protocol === "responses" ? stripResponsesText(json) : protocol === "anthropic" ? stripAnthropicText(json) : stripOpenaiChatText(json);
-      res.end(Buffer.from(JSON.stringify(stripped), "utf8"));
-      return;
-    } catch {
-    }
+  if (json && containsRenderTagText(text)) {
+    json = protocol === "responses" ? stripResponsesText(json) : protocol === "anthropic" ? stripAnthropicText(json) : stripOpenaiChatText(json);
+    mutated = true;
+  }
+  if (mutated && json) {
+    res.end(Buffer.from(JSON.stringify(json), "utf8"));
+    return;
   }
   res.end(text);
 }
@@ -56223,13 +56684,24 @@ function parseIpLiteral(s3) {
   if (t.includes(":")) return t;
   return null;
 }
+function normalizeIpLiteral(ip) {
+  const t = ip.trim().toLowerCase();
+  if (!t.startsWith("::ffff:")) return t;
+  const dotted = t.match(/^::ffff:(\d{1,3}(?:\.\d{1,3}){3})$/);
+  if (dotted) return dotted[1];
+  const hex = t.match(/^::ffff:0*([0-9a-f]{1,4}):0*([0-9a-f]{1,4})$/);
+  if (hex) {
+    const hi2 = parseInt(hex[1], 16);
+    const lo2 = parseInt(hex[2], 16);
+    return `${hi2 >> 8 & 255}.${hi2 & 255}.${lo2 >> 8 & 255}.${lo2 & 255}`;
+  }
+  return t;
+}
 function classifyIp(ip) {
-  const lit = parseIpLiteral(ip);
+  const lit = normalizeIpLiteral(parseIpLiteral(ip) ?? "");
   if (!lit) return "public";
   if (lit.includes(":")) {
     if (lit === "::1" || lit === "::") return "loopback";
-    const mapped = lit.match(/^::ffff:(\d+\.\d+\.\d+\.\d+)$/);
-    if (mapped) return classifyIp(mapped[1]);
     if (lit.startsWith("fe8") || lit.startsWith("fe9") || lit.startsWith("fea") || lit.startsWith("feb")) return "linkLocal";
     if (lit.startsWith("fc") || lit.startsWith("fd")) return "private";
     return "public";
@@ -56280,7 +56752,7 @@ async function checkTunnelDestination(origin, ctx) {
   let ips;
   const literal = parseIpLiteral(host);
   if (literal) {
-    ips = [literal];
+    ips = [normalizeIpLiteral(literal)];
   } else {
     try {
       ips = await (ctx.resolveHost ?? dnsResolveHost)(host);
@@ -56291,7 +56763,7 @@ async function checkTunnelDestination(origin, ctx) {
   }
   if (ctx.selfPort !== void 0 && port === ctx.selfPort) {
     const mine = (ctx.localIps ?? localMachineIps)();
-    if (ips.some((ip) => mine.has(ip.toLowerCase()) || mine.has(`::ffff:${ip.toLowerCase()}`))) {
+    if (ips.some((ip) => ip === "0.0.0.0" || ip === "::" || mine.has(ip.toLowerCase()) || mine.has(`::ffff:${ip.toLowerCase()}`))) {
       return { ok: false, code: "self", message: "the bili tunnel may not target the proxy itself" };
     }
   }
@@ -57168,10 +57640,7 @@ function resolveUpstream(_opts, reqUrl, req) {
   if (reqUrl.startsWith("http://") || reqUrl.startsWith("https://")) {
     try {
       const u2 = new URL(reqUrl);
-      const ownHost = (req?.headers.host ?? "").toLowerCase();
-      if (u2.host.toLowerCase() !== ownHost) {
-        return { upstream: `${u2.protocol}//${u2.host}`, rewrittenUrl: reqUrl, tunnel: true };
-      }
+      return { upstream: `${u2.protocol}//${u2.host}`, rewrittenUrl: reqUrl, tunnel: true };
     } catch {
     }
   }
@@ -57387,6 +57856,16 @@ function restoreOutputBudget(parsed, session, log2) {
     log2("info", `[${session.id}] output budget restored ${value} -> ${highWater} (#546: client shrank it from its raw-history estimate)`);
   }
 }
+var SIDE_REQUEST_GUARD_TOLERANCE = 1.15;
+function sideRequestGuard(parsed, protocol, modelContextLimit, learnedLimit) {
+  let limit = modelContextLimit;
+  if (typeof learnedLimit === "number" && learnedLimit > 0 && learnedLimit < limit) limit = learnedLimit;
+  const field = outputBudgetField(parsed);
+  const maxOut = field ? parsed[field] : 0;
+  if (limit > 0 && shouldReserveOutputHeadroom(protocol)) limit = reserveOutputHeadroom(limit, maxOut);
+  const estimate = estimateRawBodyTokens(parsed) + imageTokensInParsedBody(protocol, parsed);
+  return { blocked: limit > 0 && estimate >= limit * SIDE_REQUEST_GUARD_TOLERANCE, estimate, limit };
+}
 function isTrustedAdminOrigin(origin, host, trustedHosts) {
   if (!host || !trustedHosts.has(host.toLowerCase())) return false;
   if (!origin) return true;
@@ -57468,6 +57947,7 @@ async function handle(req, res, opts, core, config, log2, instanceId, instanceSt
       opts.proxySource = fresh.proxySource;
       opts.proxyFallback = fresh.proxyFallback;
       opts.compress = fresh.compress;
+      opts.compat = fresh.compat;
       resetProxyCache();
       for (const k2 of Object.keys(opts.routes)) delete opts.routes[k2];
       Object.assign(opts.routes, loadRoutes());
@@ -57508,7 +57988,7 @@ async function handle(req, res, opts, core, config, log2, instanceId, instanceSt
     try {
       const result = await fetchWithTimeout(targetUrl2, {
         method: "HEAD",
-        ...proxyUrl ? { dispatcher: proxyDispatcher(proxyUrl) } : {}
+        ...proxyUrl ? { dispatcher: proxyDispatcher(proxyUrl, 15e3) } : {}
       }, 15e3);
       result.clearTimer();
       recordUpstreamConnection(targetUrl2, proxyUrl);
@@ -57531,7 +58011,7 @@ async function handle(req, res, opts, core, config, log2, instanceId, instanceSt
       res.end(JSON.stringify({ ok: false, error: "conversationId query parameter is required" }));
       return;
     }
-    return handlePluginStatus(conversationId2, res, params.get("fallback") === "latest");
+    return handlePluginStatus(conversationId2, res, { core, config, log: log2 }, params.get("fallback") === "latest");
   }
   if (req.method === "POST" && req.url === "/__bili/plugin/tool") {
     try {
@@ -57787,6 +58267,25 @@ ${bodyBuffer.toString("utf8")}`);
     const pluginMode = pluginAgent !== void 0;
     restoreOutputBudget(parsed, session, log2);
     if (!countTokens && !responsesCompact && protocol !== null && isSideRequest(parsed)) {
+      const reqModel2 = parsed.model;
+      const learnedMap2 = session.metadata.learnedContextLimits;
+      const learnedLimit2 = (reqModel2 && learnedMap2 ? learnedMap2[reqModel2] : void 0) ?? session.metadata.learnedContextLimit;
+      const guard = sideRequestGuard(parsed, protocol, reqConfig.modelContextLimit, learnedLimit2);
+      if (guard.blocked) {
+        log2("warn", `[${session.id}] side request (~${guard.estimate} tokens) \u2265 effective window ${guard.limit} (model=${reqModel2 ?? "?"}) \u2014 NOT forwarded: guaranteed upstream 400 (side requests bypass preflight by design, #388)`);
+        if (!res.headersSent && !res.writableEnded && !res.destroyed) {
+          res.writeHead(413, { "content-type": "application/json" });
+          res.end(JSON.stringify({
+            error: {
+              type: "server_error",
+              code: "side_request_payload_too_large",
+              message: `side request payload ~${guard.estimate} tokens reaches the effective context window ${guard.limit} (model=${reqModel2 ?? "unknown"}); NOT forwarded \u2014 side requests (max_tokens<=${SIDE_REQUEST_MAX_TOKENS}) bypass compression by design (#388). Shrink the conversation or raise the model's context window.`,
+              retryable: false
+            }
+          }));
+        }
+        return;
+      }
       log2("info", `[${session.id}] side request (max_tokens<=${SIDE_REQUEST_MAX_TOKENS}) \u2192 passthrough + tag strip only, kernel state untouched`);
       const sidePrepared = {
         body: bodyBuffer,
@@ -57876,19 +58375,37 @@ ${bodyBuffer.toString("utf8")}`);
             instanceId
           );
           if (isPreflightFailFast(outcome)) {
-            if (outcome.respond && !res.headersSent && !res.destroyed) {
-              res.writeHead(outcome.status, {
-                "content-type": "application/json",
-                ...outcome.status === 503 ? { "retry-after": "30" } : {}
-              });
-              res.end(JSON.stringify({
-                error: {
-                  type: "server_error",
-                  code: "preflight_compress_failed",
-                  message: outcome.message,
-                  retryable: outcome.retryable
+            if (outcome.respond && !res.destroyed) {
+              if (res.headersSent) {
+                if (prepared.stream) {
+                  emitPreflightError(res, prepared.protocol, { message: outcome.message, retryable: outcome.retryable }, (m2) => log2("warn", m2));
+                } else {
+                  try {
+                    res.end(JSON.stringify({
+                      error: {
+                        type: "server_error",
+                        code: "preflight_compress_failed",
+                        message: outcome.message,
+                        retryable: outcome.retryable
+                      }
+                    }));
+                  } catch {
+                  }
                 }
-              }));
+              } else {
+                res.writeHead(outcome.status, {
+                  "content-type": "application/json",
+                  ...outcome.status === 503 ? { "retry-after": "30" } : {}
+                });
+                res.end(JSON.stringify({
+                  error: {
+                    type: "server_error",
+                    code: "preflight_compress_failed",
+                    message: outcome.message,
+                    retryable: outcome.retryable
+                  }
+                }));
+              }
             }
             return;
           }
@@ -57921,6 +58438,59 @@ function stripKernelSummaries(messages, state) {
   }
   return messages.filter((m2) => !(m2.id ?? "").startsWith("acp_summary_") || !carried.has(m2.id));
 }
+var RESPONSES_TURN_SEPARATOR = "[The exchange between these two assistant turns was compressed.]";
+function repairResponsesAssistantOrdering(folded, original) {
+  const runOf = /* @__PURE__ */ new Map();
+  const runHasBody = /* @__PURE__ */ new Map();
+  let run = 0;
+  let inRun = false;
+  for (const m2 of original) {
+    if (m2.role === "assistant") {
+      if (!inRun) {
+        run++;
+        inRun = true;
+      }
+      runOf.set(m2.id, run);
+      if (m2.contentType !== "reasoning") runHasBody.set(run, true);
+    } else {
+      inRun = false;
+    }
+  }
+  const survivorCount = /* @__PURE__ */ new Map();
+  for (const m2 of folded) {
+    const r = m2.role === "assistant" ? runOf.get(m2.id) : void 0;
+    if (r !== void 0) survivorCount.set(r, (survivorCount.get(r) ?? 0) + 1);
+  }
+  const out = [];
+  let phase = -1;
+  let seenReasoning = false;
+  let sepSeq = 0;
+  const pushSeparator = () => {
+    sepSeq++;
+    out.push({ id: `acp_turn_sep_${sepSeq}`, role: "user", contentType: "text", text: RESPONSES_TURN_SEPARATOR });
+    phase = -1;
+    seenReasoning = false;
+  };
+  for (const m2 of folded) {
+    if (m2.role !== "assistant") {
+      out.push(m2);
+      phase = -1;
+      seenReasoning = false;
+      continue;
+    }
+    const kind = m2.contentType === "reasoning" ? "reasoning" : m2.contentType === "tool-call" ? "tool-call" : "message";
+    const r = runOf.get(m2.id);
+    if (kind === "reasoning" && r !== void 0 && runHasBody.get(r) && survivorCount.get(r) === 1) continue;
+    if (kind === "reasoning" && (phase > 0 || seenReasoning) || kind === "message" && phase === 2) pushSeparator();
+    out.push(m2);
+    if (kind === "reasoning") {
+      phase = Math.max(phase, 0);
+      seenReasoning = true;
+    } else if (kind === "message") phase = Math.max(phase, 1);
+    else phase = Math.max(phase, 2);
+  }
+  return out;
+}
 function diagTagSummary(messages, sessionId, strategy) {
   let textTagged = 0;
   let toolTagged = 0;
@@ -57946,10 +58516,19 @@ function diagNudge(turn, sessionId, tokenCount, limit, model, willInject) {
   const modelTag = model ? ` model=${model}` : "";
   return `[${sessionId}] nudge ${inject}: usage=${pct2} (${tokenCount}/${limit}), growth=${growth}/${floor} (ref=${ref}, interval=${interval}), pendingT1=${pendingT1}/${interval}${modelTag}, reason="${n.reason.slice(0, 120)}"`;
 }
+function armHostUsageCredit(session, originalMessages, processedMessages, log2) {
+  session.hostCreditTokens = 0;
+  if (session.metadata.pluginAgent === "pi") return;
+  session.hostCreditTokens = processedMessages.length > 0 ? Math.max(0, estimateCoreMessages(originalMessages) - estimateCoreMessages(processedMessages)) : 0;
+  if (session.hostCreditTokens > 0) {
+    log2("info", `[${session.id}] host usage backfill armed: +${session.hostCreditTokens} tok (forwarded view is folded); host usage will report the uncompressed baseline`);
+  }
+}
 function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, session, pluginMode) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
+  session.hostCreditTokens = 0;
   const injectTools = opts.compress.injectTool && !pluginMode;
   if (isAutoModeClassifier(parsed)) {
     log2("info", `[${sessionId}] auto-mode classifier passthrough (skipping compress injection)`);
@@ -58005,20 +58584,46 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
     log2("warn", `[${sessionId}] kernel transform failed, forwarding unchanged: ${String(err2)}`);
     processedMessages = [];
   }
+  session.metadata.systemPromptTokens = countSystemAndToolsTokens(extractSystem(systemOut), toolsOut);
   snapshotMessages(session, originalMessages);
   markDirty(session);
   const rebuilt = { ...parsed, messages: rebuiltMessages, system: systemOut, tools: toolsOut };
   delete rebuilt.prompt_cache_key;
+  armHostUsageCredit(session, originalMessages, processedMessages, log2);
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, anthropicSystem: parsed.system, protocol: "anthropic", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, renderTags: "text-only" };
 }
 var OUTPUT_CLAMP_MARGIN_PCT = 0.05;
 var OUTPUT_CLAMP_MIN_MARGIN = 2048;
 var OUTPUT_CLAMP_FLOOR = 1024;
 var EMERGENCY_NUDGE_ESCALATION_PCT = 0.7;
+function countSystemAndToolsTokens(systemText, tools) {
+  return defaultCountTokens(systemText ?? "") + defaultCountTokens(JSON.stringify(tools ?? []));
+}
 function estimateInputTokens(processedMessages, systemText, tools, lastInputTokens) {
-  const est = estimateCoreMessages(processedMessages) + defaultCountTokens(systemText ?? "") + defaultCountTokens(JSON.stringify(tools ?? []));
+  const est = estimateCoreMessages(processedMessages) + countSystemAndToolsTokens(systemText, tools);
   return Math.max(lastInputTokens > 0 ? lastInputTokens : 0, est);
 }
+function estimateWireOverhead(protocol, body) {
+  let parsed;
+  try {
+    parsed = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
+  } catch {
+    return 0;
+  }
+  const sysRaw = protocol === "responses" ? parsed.instructions : parsed.system;
+  let sysText = "";
+  if (typeof sysRaw === "string") {
+    sysText = sysRaw;
+  } else if (Array.isArray(sysRaw)) {
+    sysText = sysRaw.map((part) => typeof part?.text === "string" ? part.text : "").join("\n");
+  }
+  if (protocol === "openai" && Array.isArray(parsed.messages)) {
+    const hoisted = parsed.messages.filter((m2) => m2.role === "system" || m2.role === "developer").map((m2) => typeof m2.content === "string" ? m2.content : "").join("\n");
+    sysText = sysText ? `${sysText}
+${hoisted}` : hoisted;
+  }
+  return defaultCountTokens(sysText) + defaultCountTokens(JSON.stringify(parsed.tools ?? []));
+}
 function clampOutputBudget(requested, inputEstimate, nativeWindow) {
   const margin = Math.max(OUTPUT_CLAMP_MIN_MARGIN, Math.ceil(inputEstimate * OUTPUT_CLAMP_MARGIN_PCT));
   const cap = nativeWindow - inputEstimate - margin;
@@ -58044,7 +58649,9 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
+  session.hostCreditTokens = 0;
   let openaiSystemText = "";
+  let openaiOutboundSystem;
   let processedMessages = [];
   let originalMessages = [];
   let nudge;
@@ -58085,6 +58692,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
     if (systemText) sysParts.push(systemText);
     if (shouldInject) sysParts.push(buildCompressSystemPrompt(prompts));
     rebuiltMessages = injectOpenaiSystem(rebuiltMessages, sysParts);
+    openaiOutboundSystem = sysParts.join("\n\n");
     if (injectTools) {
       toolsOut = injectOpenaiTool(parsed.tools);
     }
@@ -58107,6 +58715,10 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
   if (stream2 && rebuilt.stream_options === void 0) {
     rebuilt.stream_options = { include_usage: true };
   }
+  armHostUsageCredit(session, originalMessages, processedMessages, log2);
+  if (!isTitleGen && openaiOutboundSystem !== void 0) {
+    session.metadata.systemPromptTokens = countSystemAndToolsTokens(openaiOutboundSystem, toolsOut);
+  }
   snapshotMessages(session, originalMessages);
   markDirty(session);
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, protocol: "openai", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, openaiSystemText, renderTags: "text-only" };
@@ -58115,6 +58727,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
+  session.hostCreditTokens = 0;
   if (reconcileNativeCompactionBoundary(session)) {
     log2("info", `[${sessionId}] reconciled ACP state after native Responses compact boundary`);
   }
@@ -58134,6 +58747,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
   let rebuiltInput = parsed.input;
   let toolsOut = parsed.tools;
   let transformOk = false;
+  let responsesDevContent;
   const typedItems = normalizeResponsesMessageItems(parsed.input);
   if (typedItems > 0) {
     log2("info", `[${sessionId}] stamped type:"message" on ${typedItems} type-less input item(s) before projection (omp wire form)`);
@@ -58174,19 +58788,21 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
     const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && !isCompactionTrigger && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
-    processedMessages = stripKernelSummaries(turn.messages, turn.state);
+    processedMessages = repairResponsesAssistantOrdering(stripKernelSummaries(turn.messages, turn.state), originalMessages);
     reapOrphanBlocks(session, msgs, deactivateBlock);
     rebuiltInput = patchResponsesInput(projection, processedMessages);
     const forgedSummaries = echoReplaced ? [] : session.metadata.codexForgedSummaries ?? [];
     if (shouldInject && !isCompactionTrigger && !process.env.ACP_NO_COMPRESS_PROMPT) {
       const prompt = responsesTextProtocol ? buildCompressHybridSystemPrompt(prompts) : buildCompressSystemPrompt(prompts);
       const devContent = [...projection.systemParts, ...forgedSummaries, prompt].join("\n\n---\n\n");
+      responsesDevContent = devContent;
       rebuiltInput = injectResponsesDeveloperMessage(rebuiltInput, devContent);
       if (!process.env.ACP_NO_INJECT_TOOL && injectTools) {
         toolsOut = responsesTextProtocol ? injectResponsesTool(parsed.tools, ACP_READONLY_TOOLS_RESPONSES) : injectResponsesTool(parsed.tools);
       }
     } else if (projection.systemParts.length > 0 || forgedSummaries.length > 0) {
       const devContent = [...projection.systemParts, ...forgedSummaries].join("\n\n---\n\n");
+      responsesDevContent = devContent;
       rebuiltInput = injectResponsesDeveloperMessage(rebuiltInput, devContent);
     }
     if (willInjectNudge && turn.nudge) {
@@ -58244,6 +58860,10 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     });
     log2("info", `[${sessionId}] responses forward tools=[${fwdTools.join(",")}] injectTool=${injectTools}${pluginMode ? " (plugin mode: wire injection suppressed)" : ""} NO_INJECT_TOOL=${!!process.env.ACP_NO_INJECT_TOOL} NO_COMPRESS_PROMPT=${!!process.env.ACP_NO_COMPRESS_PROMPT}`);
   }
+  armHostUsageCredit(session, originalMessages, processedMessages, log2);
+  if (transformOk) {
+    session.metadata.systemPromptTokens = countSystemAndToolsTokens(responsesDevContent ?? "", toolsOut);
+  }
   snapshotMessages(session, originalMessages);
   markDirty(session);
   return {
@@ -58330,7 +58950,7 @@ function prepareResponsesCompact(body, parsed, session, req, core, config, log2)
       session.state = prevState;
       return base;
     }
-    const processed = stripKernelSummaries(turn.messages, turn.state);
+    const processed = repairResponsesAssistantOrdering(stripKernelSummaries(turn.messages, turn.state), projection.msgs);
     const output = patchResponsesInput(projection, processed);
     if (typeof output === "string") {
       session.state = prevState;
@@ -58415,6 +59035,12 @@ function logUpstreamProxyDecision(opts, upstreamUrl, decision) {
   const via = decision.proxy ? `via ${maskUrlForLog(decision.proxy)}` : "direct";
   logMsg(opts, "info", `[upstream-proxy] ${maskHostPortForLog(host)} ${via} (source=${decision.source})`);
 }
+function inferWireProtocol(path18) {
+  const p2 = path18.split("?", 2)[0];
+  if (p2.endsWith("/chat/completions")) return "openai";
+  if (p2.endsWith("/responses") || p2.endsWith("/responses/compact")) return "responses";
+  return null;
+}
 function buildForwardTarget(req, opts, route, affinity, hopMarker) {
   const reqUrl = req.url ?? "";
   const isAbsoluteUrl = /^https?:\/\//i.test(reqUrl);
@@ -58446,12 +59072,53 @@ function buildForwardTarget(req, opts, route, affinity, hopMarker) {
 function isPreflightFailFast(outcome) {
   return "failFast" in outcome;
 }
+var PREFLIGHT_HOLD_GRACE_DEFAULT_MS = 3e4;
+var PREFLIGHT_KEEPALIVE_MS = 15e3;
+function preflightHoldGraceMs() {
+  const raw = process.env.BILI_PREFLIGHT_HOLD_MS;
+  if (!raw) return PREFLIGHT_HOLD_GRACE_DEFAULT_MS;
+  const v2 = Number(raw);
+  return Number.isFinite(v2) && v2 >= 0 ? Math.floor(v2) : PREFLIGHT_HOLD_GRACE_DEFAULT_MS;
+}
+function beginPreflightHold(res, prepared, log2) {
+  if (res.headersSent || res.destroyed || res.writableEnded) return void 0;
+  const sid = prepared.session.id;
+  const keepAlive = prepared.stream ? ": bili-preflight\n\n" : " ";
+  try {
+    if (prepared.stream) {
+      res.writeHead(200, {
+        "content-type": "text/event-stream",
+        "cache-control": "no-cache",
+        "x-accel-buffering": "no",
+        "x-bili-preflight": "compressing"
+      });
+    } else {
+      res.writeHead(200, { "content-type": "application/json", "x-bili-preflight": "compressing" });
+    }
+  } catch {
+    return void 0;
+  }
+  log2("info", `[${sid}] preflight still running after ${preflightHoldGraceMs()}ms grace \u2014 committed early ${prepared.stream ? "SSE" : "JSON"} headers + keep-alive to hold the client (#568)`);
+  try {
+    res.write(keepAlive);
+  } catch {
+  }
+  const iv = setInterval(() => {
+    try {
+      res.write(keepAlive);
+    } catch {
+      clearInterval(iv);
+    }
+  }, PREFLIGHT_KEEPALIVE_MS);
+  return () => clearInterval(iv);
+}
 async function preflightCompressIfNeeded(prepared, runPrepare, req, res, opts, core, config, model, route, affinity, log2, instanceId) {
   const session = prepared.session;
   const limit = config.modelContextLimit;
   const imageTokens = imageTokensInRawBody(prepared.protocol, prepared.body);
   const textEstimate = estimateCoreMessages(prepared.processedMessages);
-  const payloadEstimate = textEstimate + imageTokens;
+  const overheadEstimate = estimateWireOverhead(prepared.protocol, prepared.body);
+  const payloadEstimate = textEstimate + overheadEstimate + imageTokens;
   const tokenCount = Math.max(session.stats.lastInputTokens, payloadEstimate);
   if (limit <= 0 || !model || tokenCount < limit) return prepared;
   const learnedMap = session.metadata.learnedContextLimits;
@@ -58467,12 +59134,9 @@ async function preflightCompressIfNeeded(prepared, runPrepare, req, res, opts, c
     log2("error", `[${session.id}] preflight fail-fast ${status2} (retryable=${retryable}): ${message}`);
     return { failFast: true, status: status2, message, retryable, respond: !res.writableEnded };
   };
-  if ((prepared.nudge?.compressibleRanges ?? []).length === 0) {
-    if (payloadEstimate < limit) {
-      log2("warn", `[${session.id}] preflight trigger fired on a stale baseline (~${tokenCount}) but the payload fits (~${payloadEstimate}/${limit}); forwarding as-is`);
-      return prepared;
-    }
-    return failFast(502, "no part of the conversation is compressible (nothing left to fold)", false);
+  if ((prepared.nudge?.compressibleRanges ?? []).length === 0 && payloadEstimate < limit) {
+    log2("warn", `[${session.id}] preflight trigger fired on a stale baseline (~${tokenCount}) but the payload fits (~${payloadEstimate}/${limit}); forwarding as-is`);
+    return prepared;
   }
   log2("warn", `[${session.id}] context ${tokenCount} tokens exceeds model window ${limit} (model=${model}); preflight compressing before forward`);
   const { upstreamUrl, headers, proxyUrl } = buildForwardTarget(req, opts, route, affinity, instanceId);
@@ -58481,31 +59145,42 @@ async function preflightCompressIfNeeded(prepared, runPrepare, req, res, opts, c
     if (!res.writableEnded) clientAbort.abort();
   });
   const started = Date.now();
-  const result = await preflightCompress(
-    {
-      core,
-      session,
-      config,
-      prompts: prepared.prompts ?? defaultPrompts,
-      protocol: prepared.protocol,
-      url: upstreamUrl,
-      headers,
-      model,
-      proxyUrl,
-      signal: clientAbort.signal,
-      log: log2,
-      imageFloor: imageTokens
-    },
-    prepared.originalMessages
-  );
-  if (result.payloadEstimate < limit) {
-    if (result.compressedRanges > 0) {
-      log2("info", `[${session.id}] preflight compressed ${result.compressedRanges} range(s), ~${result.savedTokens} tokens saved (${tokenCount} \u2192 ${session.stats.lastInputTokens}) in ${Date.now() - started}ms; rebuilding payload`);
-      const rebuilt = runPrepare();
-      session.stats.requests -= 1;
-      return rebuilt;
-    }
-    log2("warn", `[${session.id}] preflight made no progress but the payload fits (~${result.payloadEstimate}/${limit}); forwarding as-is`);
+  let stopHold;
+  const holdTimer = setTimeout(() => {
+    stopHold = beginPreflightHold(res, prepared, log2);
+  }, preflightHoldGraceMs());
+  holdTimer.unref();
+  let result;
+  try {
+    result = await preflightCompress(
+      {
+        core,
+        session,
+        config,
+        prompts: prepared.prompts ?? defaultPrompts,
+        protocol: prepared.protocol,
+        url: upstreamUrl,
+        headers,
+        model,
+        proxyUrl,
+        signal: clientAbort.signal,
+        log: log2,
+        imageFloor: imageTokens,
+        wireOverhead: overheadEstimate
+      },
+      prepared.originalMessages
+    );
+  } finally {
+    clearTimeout(holdTimer);
+    stopHold?.();
+  }
+  if (result.compressedRanges > 0) {
+    log2("info", `[${session.id}] preflight compressed ${result.compressedRanges} range(s), ~${result.savedTokens} tokens saved (${tokenCount} \u2192 ${session.stats.lastInputTokens}) in ${Date.now() - started}ms; rebuilding payload`);
+    const rebuilt = runPrepare();
+    session.stats.requests -= 1;
+    if (estimateCoreMessages(rebuilt.processedMessages) + overheadEstimate + imageTokens < limit) return rebuilt;
+  } else if (estimateCoreMessages(prepared.processedMessages) + overheadEstimate + imageTokens < limit) {
+    log2("warn", `[${session.id}] preflight made no progress but the payload fits; forwarding as-is`);
     return prepared;
   }
   const f2 = result.failure;
@@ -58514,16 +59189,40 @@ async function preflightCompressIfNeeded(prepared, runPrepare, req, res, opts, c
     return { failFast: true, status: 0, message: f2.detail, retryable: false, respond: false };
   }
   const status = f2?.kind === "upstream" && f2.status === 429 ? 503 : 502;
-  return failFast(status, f2?.detail ?? "unknown reason", status === 503);
+  return failFast(status, f2?.detail ?? "the payload still exceeds the window after preflight compression", status === 503);
 }
 async function forward(req, res, opts, body, prepared, core, config, log2, route, instanceId, affinity) {
   if (prepared?.codexForge) {
     log2("info", `[${prepared.session.id}] codex compact served locally (${prepared.codexForge.kind}); upstream not contacted`);
-    res.writeHead(200, { "content-type": prepared.codexForge.contentType });
+    if (!res.headersSent) res.writeHead(200, { "content-type": prepared.codexForge.contentType });
     res.end(prepared.codexForge.body);
     return;
   }
+  let wireBody = body;
+  let compatRoles = null;
+  let compatProtocol = null;
   const { upstreamUrl, headers, proxyUrl } = buildForwardTarget(req, opts, route, affinity, prepared !== null ? instanceId : void 0);
+  if (typeof body === "string") {
+    const configured = resolveCompatRoles(opts.routes, upstreamUrl, opts.compat?.roles);
+    const learned = prepared?.session.metadata.learnedCompatRoles ?? {};
+    const roles = { ...configured, ...learned };
+    const protocol = prepared?.protocol ?? route?.explicitProtocol ?? inferWireProtocol(req.url ?? "");
+    if (protocol === "openai" || protocol === "responses") {
+      compatProtocol = protocol;
+      if (Object.keys(roles).length > 0) {
+        compatRoles = roles;
+        const applied = applyCompatRoles(body, protocol, roles);
+        if (applied.rewritten > 0) {
+          wireBody = applied.body;
+          log2("info", `[${prepared?.session.id ?? "passthrough"}] [compat] rewrote ${applied.rewritten} message role(s) per compat.roles (${Object.entries(roles).map(([f2, t]) => `${f2}\u2192${t}`).join(",")})`);
+        }
+      }
+    }
+  }
+  const wireTransform = compatProtocol ? (b2) => {
+    if (compatRoles) applyCompatRolesJson(b2, compatProtocol, compatRoles);
+    return b2;
+  } : void 0;
   log2("info", `forward ${req.method} \u2192 ${maskUrlForLog(upstreamUrl)}`);
   if (process.env.ACP_DEBUG && prepared) {
     const sid = prepared.session.id;
@@ -58538,9 +59237,9 @@ async function forward(req, res, opts, body, prepared, core, config, log2, route
       }
     }
   }
-  if (typeof body === "string" && (opts.debug || bodyDumpEnabled())) {
+  if (typeof wireBody === "string" && (opts.debug || bodyDumpEnabled())) {
     try {
-      const parsed = JSON.parse(body);
+      const parsed = JSON.parse(wireBody);
       if (opts.debug) {
         const toolNames = (parsed.tools ?? []).map((t) => {
           const fn = t.function;
@@ -58557,10 +59256,10 @@ async function forward(req, res, opts, body, prepared, core, config, log2, route
         const sid = prepared?.session.id ?? "unknown";
         const out = `${dumpDir}/req-${Date.now()}-${safeSessionId(sid)}.json`;
         try {
-          const pretty = JSON.stringify(JSON.parse(body), null, 2);
+          const pretty = JSON.stringify(JSON.parse(wireBody), null, 2);
           fs8.writeFileSync(out, pretty);
         } catch {
-          fs8.writeFileSync(out, body);
+          fs8.writeFileSync(out, wireBody);
         }
         log2("info", `[debug] forwarded body written to ${out}`);
       }
@@ -58589,7 +59288,7 @@ async function forward(req, res, opts, body, prepared, core, config, log2, route
   if (rawBase) {
     try {
       const hdrText = Object.entries(maskHeadersForLog(headers)).map(([k2, v2]) => `${k2}: ${v2}`).join("\n");
-      const bodyText = req.method === "GET" || req.method === "HEAD" ? "" : typeof body === "string" ? body : Buffer.from(body).toString("utf8");
+      const bodyText = req.method === "GET" || req.method === "HEAD" ? "" : typeof wireBody === "string" ? wireBody : Buffer.from(wireBody).toString("utf8");
       const reqPath = `${rawBase}-REQ.txt`;
       fs8.writeFileSync(reqPath, `${req.method ?? "POST"} ${maskUrlForLog(upstreamUrl)}
 ${hdrText}
@@ -58604,7 +59303,7 @@ ${bodyText}`);
   const init = {
     method: req.method ?? "GET",
     headers,
-    body: req.method === "GET" || req.method === "HEAD" ? void 0 : body
+    body: req.method === "GET" || req.method === "HEAD" ? void 0 : wireBody
   };
   if (dispatcher) init.dispatcher = dispatcher;
   const clientAbort = new AbortController();
@@ -58619,6 +59318,49 @@ ${bodyText}`);
     recordUpstreamConnection(upstreamUrl, proxyUrl, error);
     throw new Error(`upstream request failed: ${formatUpstreamError(error, upstreamUrl, proxyUrl)}`, { cause: error });
   }
+  if (compatProtocol && typeof wireBody === "string" && upstreamResult.response.status === 400 && upstreamResult.response.body) {
+    let roleErrText = null;
+    try {
+      roleErrText = (await readStreamToBuffer(upstreamResult.response.body)).toString("utf8");
+    } catch {
+      roleErrText = null;
+    }
+    if (roleErrText !== null) {
+      upstreamResult = {
+        response: new Response(roleErrText, {
+          status: upstreamResult.response.status,
+          statusText: upstreamResult.response.statusText,
+          headers: new Headers(upstreamResult.response.headers)
+        }),
+        clearTimer: upstreamResult.clearTimer
+      };
+      const rejection = detectRoleRejection(upstreamResult.response.status, roleErrText);
+      if (rejection && rejection.role !== "system") {
+        const fixed = applyCompatRoles(wireBody, compatProtocol, { [rejection.role]: "system" });
+        if (fixed.rewritten > 0) {
+          try {
+            const retry = await fetchWithTimeout(upstreamUrl, { ...init, body: fixed.body }, void 0, clientAbort.signal);
+            if (retry.response.ok) {
+              upstreamResult.clearTimer();
+              const s3 = prepared?.session;
+              if (s3) {
+                const prev = s3.metadata.learnedCompatRoles ?? {};
+                s3.metadata.learnedCompatRoles = { ...prev, [rejection.role]: "system" };
+                markDirty(s3);
+              }
+              compatRoles = { ...compatRoles, [rejection.role]: "system" };
+              const providerKey = new URL(upstreamUrl).origin;
+              log2("info", `[${prepared?.session.id ?? "passthrough"}] [compat] upstream rejected role "${rejection.role}" \u2014 auto-rewrote ${fixed.rewritten} message role(s) to "system", retry OK (remembered for this session only). To make permanent, add: {"providers":{"${providerKey}":{"compat":{"roles":{"${rejection.role}":"system"}}}}`);
+              upstreamResult = retry;
+            } else {
+              retry.clearTimer();
+            }
+          } catch {
+          }
+        }
+      }
+    }
+  }
   const { response: upstream, clearTimer: clearUpstreamTimer } = upstreamResult;
   const respHeaders = {};
   const respConnNamed = connectionNamedHeaders(upstream.headers.get("connection") ?? void 0);
@@ -58664,9 +59406,10 @@ ${hdrText}
       if (info.isOverflow) {
         let reqModel;
         let rejectedImageTokens = 0;
+        let parsedBody;
         try {
           const rawBody = typeof prepared.body === "string" ? prepared.body : prepared.body.toString("utf8");
-          const parsedBody = JSON.parse(rawBody);
+          parsedBody = JSON.parse(rawBody);
           reqModel = typeof parsedBody.model === "string" ? parsedBody.model : void 0;
           rejectedImageTokens = imageTokensInParsedBody(prepared.protocol, parsedBody);
         } catch {
@@ -58680,7 +59423,7 @@ ${hdrText}
           s3.metadata.learnedContextLimits = learnedMap;
           log2("warn", `[${s3.id}] upstream context overflow \u2014 learned real window ${info.window} for ${reqModel ?? "(unknown model)"} (was ${prev ?? "unset"}); arming emergency shrink`);
         } else {
-          const payloadEstimate = estimateCoreMessages(prepared.processedMessages) + rejectedImageTokens;
+          const payloadEstimate = (prepared.processedMessages.length > 0 ? estimateCoreMessages(prepared.processedMessages) : estimateRawBodyTokens(parsedBody)) + rejectedImageTokens;
           const prev = (reqModel ? learnedMap[reqModel] : void 0) ?? s3.metadata.learnedContextLimit;
           if (payloadEstimate >= 1e3 && (prev === void 0 || payloadEstimate < prev)) {
             if (reqModel) learnedMap[reqModel] = payloadEstimate;
@@ -58704,6 +59447,18 @@ ${hdrText}
     if (bodyText.length > 600) snippet += " \u2026";
     if (!snippet) snippet = "(no body)";
     log("warn", `[${errSid}] \u2190 upstream ${upstream.status}${reqIdText}: ${snippet}`);
+    if (res.headersSent) {
+      if (prepared?.stream) {
+        emitStreamError(res, prepared.protocol, `upstream HTTP ${upstream.status}: ${snippet}`, (m2) => log("info", m2));
+      } else {
+        try {
+          res.end(errBody ?? void 0);
+        } catch {
+        }
+      }
+      clearUpstreamTimer();
+      return;
+    }
     const errHeaders = { ...respHeaders };
     delete errHeaders["content-length"];
     delete errHeaders["transfer-encoding"];
@@ -58712,7 +59467,7 @@ ${hdrText}
     clearUpstreamTimer();
     return;
   }
-  res.writeHead(upstream.status, respHeaders);
+  if (!res.headersSent) res.writeHead(upstream.status, respHeaders);
   if (!upstream.body) {
     res.end();
     clearUpstreamTimer();
@@ -58820,7 +59575,7 @@ ${hdrText}
         const reqHeaders = buildForwardHeaders(headers);
         const textProtocol = prepared.protocol === "responses" && !!prepared.responsesTextProtocol;
         const systemPrompt = textProtocol ? buildCompressHybridSystemPrompt(prepared.prompts ?? defaultPrompts) : buildCompressSystemPrompt(prepared.prompts ?? defaultPrompts);
-        const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText);
+        const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText, prepared.session.hostCreditTokens ?? 0);
         const refreshFolded = (current) => {
           const turn = core.processTurn({
             messages: prepared.originalMessages,
@@ -58831,13 +59586,13 @@ ${hdrText}
           });
           prepared.session.state = turn.state;
           const records = current.filter((m2) => typeof m2.id === "string" && m2.id.startsWith("acp_loop_"));
-          return stripKernelSummaries([...turn.messages, ...records], turn.state);
+          return repairResponsesAssistantOrdering(stripKernelSummaries([...turn.messages, ...records], turn.state), prepared.originalMessages);
         };
         const loop = runCompressLoop(
           streamToRead,
-          { core, config, messages: prepared.processedMessages.length > 0 ? prepared.processedMessages : prepared.originalMessages, compressMessages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, protocol: prepared.protocol, textProtocol, debug: opts.debug, nudge: prepared.nudge, refreshFolded },
+          { core, config, messages: prepared.processedMessages.length > 0 ? prepared.processedMessages : prepared.originalMessages, compressMessages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, protocol: prepared.protocol, textProtocol, debug: opts.debug, refreshFolded },
           parsedReq,
-          { url: upstreamUrl, headers: reqHeaders },
+          { url: upstreamUrl, headers: reqHeaders, wireTransform },
           adapter,
           systemPrompt,
           clientAbort.signal
@@ -58884,7 +59639,7 @@ ${hdrText}
               json,
               { core, config, messages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, textProtocol: true },
               requestBody,
-              { url: upstreamUrl, headers: requestHeaders }
+              { url: upstreamUrl, headers: requestHeaders, wireTransform }
             );
           }
           const u2 = json.usage ?? {};
@@ -58902,6 +59657,10 @@ ${hdrText}
             const out = u2.completion_tokens ?? u2.output_tokens;
             if (typeof out === "number") prepared.session.stats.outputTokens += out;
           }
+          const credit = prepared.session.hostCreditTokens ?? 0;
+          if (credit > 0 && backfillHostUsage(prepared.protocol, u2, credit)) {
+            prepared.session.hostContextTokens = (typeof total === "number" ? total : 0) + credit;
+          }
           if (prepared.protocol === "openai") {
             rewriteOpenaiJsonResponse(json, ctx);
           } else if (prepared.protocol === "responses") {
@@ -59045,6 +59804,7 @@ function handleConfigReload(opts, res, log2) {
   for (const k2 of Object.keys(opts.routes)) delete opts.routes[k2];
   Object.assign(opts.routes, fresh);
   opts.compress = loadOptions().compress;
+  opts.compat = loadOptions().compat;
   resetProxyCache();
   const names = Object.keys(fresh);
   log2("info", `[acp-web] routes hot-reloaded (${names.length} providers): ${names.join(", ") || "(none)"}`);
@@ -62148,6 +62908,14 @@ async function findInstallDir(packageName) {
     dir = parent;
   }
 }
+async function isGitWorkingTree(dir) {
+  try {
+    await access(path12.join(dir, ".git"));
+    return true;
+  } catch {
+    return false;
+  }
+}
 async function readDiskVersion(installDir) {
   try {
     const pkg = JSON.parse(await readFile2(path12.join(installDir, "package.json"), "utf-8"));
@@ -62304,6 +63072,11 @@ async function checkForUpdate(opts, force = false) {
     }
     await writeLastCheck(now);
     firstCheckDone = true;
+    const installDir = await findInstallDir(opts.packageName);
+    if (installDir && await isGitWorkingTree(installDir)) {
+      log("info", `[update] running from a source checkout (${installDir}) \u2014 skipping auto-update (use npm install -g ${opts.packageName})`);
+      return;
+    }
     log("info", `[update] checking npm registry for ${opts.packageName}${sinceLastSec < 0 ? " (startup check)" : sinceLastSec === 0 ? "" : ` (last check ${sinceLastSec}s ago)`}\u2026`);
     const url = `${REGISTRY_BASE}/${opts.packageName}/latest`;
     const res = await fetch(url, {
@@ -62320,7 +63093,6 @@ async function checkForUpdate(opts, force = false) {
       log("warn", `[update] registry response had no version, skipping`);
       return;
     }
-    const installDir = await findInstallDir(opts.packageName);
     const diskVersion = installDir ? await readDiskVersion(installDir) : void 0;
     const currentVersion = diskVersion ?? opts.currentVersion;
     if (!isNewer(latest, currentVersion)) {
@@ -62391,6 +63163,9 @@ async function installViaTarball(version2, tarballUrl, installDir, integrity, sh
   } catch {
     return { ok: false, error: `install dir not writable: ${installDir}` };
   }
+  if (await isGitWorkingTree(installDir)) {
+    return { ok: false, error: `install dir is a git working tree (${installDir}) \u2014 refusing to overwrite a source checkout (use npm install -g)` };
+  }
   const MAX_TARBALL_BYTES = 100 * 1024 * 1024;
   let tgzBuffer;
   try {
@@ -63104,6 +63879,11 @@ function unwrapUpstream2(url) {
   const idx = url.indexOf("/bili/");
   return idx >= 0 ? url.slice(idx + "/bili/".length) : url;
 }
+function isLoopbackHost(host) {
+  const h = host.toLowerCase();
+  if (h === "localhost" || h === "::1" || h === "[::1]") return true;
+  return /^127\.\d+\.\d+\.\d+$/.test(h);
+}
 function resolveCaCertPath(env) {
   const base = env.XDG_DATA_HOME || path15.join(os5.homedir(), ".local/share");
   return path15.join(base, "billion-context", "ca", "root-ca.pem");
@@ -63116,6 +63896,7 @@ function discoverRoutes(client, config) {
   const httpsDomains = [];
   const httpRewrites = [];
   const httpsRewrites = [];
+  const httpEnvRoutes = [];
   const httpsSeen = /* @__PURE__ */ new Set();
   const rewriteKeys = /* @__PURE__ */ new Set();
   const httpsRewriteKeys = /* @__PURE__ */ new Set();
@@ -63201,8 +63982,18 @@ function discoverRoutes(client, config) {
         if (dshSeen.has(real)) continue;
         dshSeen.add(real);
         anon += 1;
-        rewriteKeys.add(`dsh-${anon}`);
-        httpRewrites.push({ key: `dsh-${anon}`, realUpstream: real });
+        if (isLoopbackHost(url.hostname)) {
+          rewriteKeys.add(`dsh-${anon}`);
+          httpRewrites.push({ key: `dsh-${anon}`, realUpstream: real });
+        } else if (url.protocol === "https:") {
+          const host = url.hostname.toLowerCase();
+          if (host && !httpsSeen.has(host)) {
+            httpsSeen.add(host);
+            httpsDomains.push(host);
+          }
+        } else if (!httpEnvRoutes.includes(real)) {
+          httpEnvRoutes.push(real);
+        }
       } catch {
       }
     }
@@ -63212,7 +64003,7 @@ function discoverRoutes(client, config) {
     }
     classify(config.codex?.openaiBaseUrl, "openai_base_url");
   }
-  return { httpsDomains, httpRewrites, httpsRewrites };
+  return { httpsDomains, httpRewrites, httpsRewrites, httpEnvRoutes };
 }
 function discoverDomains(client, config) {
   return discoverRoutes(client, config).httpsDomains;
@@ -64116,7 +64907,7 @@ async function runLaunch(params, deps = {}) {
   const domains = dedupeInOrder([...routes.httpsDomains, ...params.mitmDomains ?? []]);
   const handle2 = await ensureProxyRunning({ host, port, passthrough, debug, mitmDomains: domains, modelWindows: collectModelWindows(config, base) }, deps);
   console.error(
-    `bili: started proxy at ${handle2.origin} (MITM domains: ${domains.length ? domains.join(", ") : "defaults"})` + (routes.httpRewrites.length > 0 ? ` (HTTP /bili/ rewrites: ${routes.httpRewrites.length})` : "") + (routes.httpsRewrites.length > 0 ? ` (HTTPS cert rewrites: ${routes.httpsRewrites.length})` : "") + (params.client === "pi-test" ? " (no extensions)" : "")
+    `bili: started proxy at ${handle2.origin} (MITM domains: ${domains.length ? domains.join(", ") : "defaults"})` + (routes.httpRewrites.length > 0 ? ` (HTTP /bili/ rewrites: ${routes.httpRewrites.length})` : "") + (routes.httpsRewrites.length > 0 ? ` (HTTPS cert rewrites: ${routes.httpsRewrites.length})` : "") + (routes.httpEnvRoutes.length > 0 ? ` (HTTP proxy-env routes: ${routes.httpEnvRoutes.length})` : "") + (params.client === "pi-test" ? " (no extensions)" : "")
   );
   if (handle2.logPath) {
     console.error(`bili: proxy log: ${handle2.logPath}`);
@@ -64181,18 +64972,25 @@ async function runLaunch(params, deps = {}) {
       );
     }
   } else if (base === "dsh") {
-    env = { ...process.env, BILLION_CONTEXT_PROXY: origin };
+    const usesProxyEnv = routes.httpsDomains.length > 0 || routes.httpEnvRoutes.length > 0;
+    env = usesProxyEnv ? stripInheritedProxy(process.env) : { ...process.env };
+    env.BILLION_CONTEXT_PROXY = origin;
     env.DEEPSEEK_BASE_URL = wrapUpstream(origin, "https://api.deepseek.com");
+    if (usesProxyEnv) {
+      env.HTTPS_PROXY = origin;
+      env.SSL_CERT_FILE = resolveCombinedCaPath(process.env);
+    }
+    if (routes.httpEnvRoutes.length > 0) env.HTTP_PROXY = origin;
     env.PI_CACHE_RETENTION = "long";
     const dshHomeDir = resolveDshHome(process.env);
-    dshOverlayHome = prepareDshHome(dshHomeDir, origin, routes.httpRewrites);
+    dshOverlayHome = routes.httpRewrites.length > 0 ? prepareDshHome(dshHomeDir, origin, routes.httpRewrites) : void 0;
     if (dshOverlayHome) {
       env.DSH_HOME = dshOverlayHome;
     } else if (routes.httpRewrites.length > 0) {
       console.error(
-        "bili: dsh settings.yaml could not be rewritten (unreadable or no matching endpoints) \u2014 custom providers will NOT go through the proxy; the built-in deepseek route still does."
+        "bili: dsh settings.yaml could not be rewritten (unreadable or no matching endpoints) \u2014 loopback custom providers will NOT go through the proxy; other routes still do."
       );
-    } else {
+    } else if (!usesProxyEnv) {
       console.error(
         "bili: no custom providers found in ~/.dsh/settings.yaml \u2014 proxying the built-in deepseek route via DEEPSEEK_BASE_URL only."
       );
@@ -64355,10 +65153,13 @@ function renderHandoff(s3, full) {
   lines.push("");
   const messages = s3.lastMessages;
   if (messages && messages.length > 0) {
-    lines.push(full ? `## Full conversation (${messages.length} messages)` : `## Conversation (folded view as the model saw it, ${messages.length} client messages)`);
+    const folded = s3.lastMessagesFolded === true;
+    const view = full || folded ? messages : prune(messages, s3.state);
+    lines.push(
+      full && !folded ? `## Full conversation (${messages.length} messages)` : folded ? `## Conversation (persisted folded snapshot, ${messages.length} messages)` : `## Conversation (folded view as the model saw it, ${messages.length} client messages)`
+    );
     lines.push("");
     let lastRole = "";
-    const view = full ? messages : prune(messages, s3.state);
     for (const m2 of view) {
       if (m2.role !== lastRole) {
         lines.push(`### ${m2.role}`);
@@ -64368,6 +65169,18 @@ function renderHandoff(s3, full) {
       lines.push(renderMessage2(m2));
     }
     lines.push("");
+    if (full && folded) {
+      for (const b2 of s3.state.blocks.filter((x) => x.active)) {
+        const content = s3.blockContents.get(b2.blockId);
+        if (!content) continue;
+        lines.push(`## Block ${b2.blockId}${b2.topic ? ` \u2014 ${b2.topic}` : ""}`);
+        lines.push("");
+        lines.push(`### Original messages (${content.full.count})`);
+        lines.push("");
+        lines.push(content.full.text.trim());
+        lines.push("");
+      }
+    }
     return lines.join("\n");
   }
   const active = s3.state.blocks.filter((b2) => b2.active);
@@ -64429,7 +65242,7 @@ async function exportSession(selector, opts = {}) {
   const store = new SessionStore({ dir: opts.dir, enabled: true });
   const all = [...(await store.loadAll()).values()];
   if (all.length === 0) {
-    return "No persisted sessions found. Sessions are written under the sessions directory once the proxy has served a request (compression state and compressed originals only \u2014 uncompressed conversation text is not persisted).";
+    return "No persisted sessions found. Sessions are written under the sessions directory once the proxy has served a request (compression state, compressed originals, and a bounded folded-view snapshot of the recent conversation).";
   }
   if (!selector) {
     const list = await listSessions2(opts);