@prismshadow/penguin-server 0.2.16 → 0.2.17

package/dist/chunk-GGIINEEK.js

@@ -0,0 +1,3401 @@
+// src/model-catalog/catalog.ts
+import {
+  CLIENT_CREDENTIALS,
+  OFFICIAL_CLIENT_TYPES,
+  routedClientType
+} from "@prismshadow/mmsp/credentials";
+import { sameClientType } from "@prismshadow/penguin-core/model-connection";
+var VLLM_CLIENT_TYPE = "openai-chat-vllm-adapter";
+var VLLM_LOCAL_BASE_URL = "http://127.0.0.1:8000/v1";
+var OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1";
+var SILICONFLOW_BASE_URL = "https://api.siliconflow.cn/v1";
+var QWEN_TOKEN_PLAN_BASE_URL = "https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1";
+var QWEN_PAYG_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1";
+var FIREWORKS_BASE_URL = "https://api.fireworks.ai/inference/v1";
+var TOKENDANCE_BASE_URL = "https://tokendance.space/gateway/v1";
+var TOKENDANCE_MESSAGES_BASE_URL = "https://tokendance.space/gateway";
+var DEEPSEEK_BASE_URL = "https://api.deepseek.com";
+var MODELSCOPE_BASE_URL = "https://api-inference.modelscope.cn/v1";
+var ATLASCLOUD_BASE_URL = "https://api.atlascloud.ai/v1";
+var PENGUIN_GO_BASE_URL = "https://go.penguin.ooo/api/v1";
+var OPENCODE_GO_BASE_URL = "https://opencode.ai/zen/go/v1";
+var OPENCODE_GO_MESSAGES_BASE_URL = "https://opencode.ai/zen/go";
+var OPENCODE_GO_GOOGLE_BASE_URL = "https://opencode.ai/inference/go/google";
+var PENGUIN_GO_PROVIDER_ID = "penguin-go";
+function fixedClientType(providerId) {
+  return providerId === PENGUIN_GO_PROVIDER_ID ? "mmsp" : void 0;
+}
+function fixedClientTypeRefuses(providerId, target, next, stored) {
+  const fixed = fixedClientType(providerId);
+  if (fixed === void 0) return false;
+  const value = next?.trim() ?? "";
+  if (value === "") return target === "group" && (stored?.trim() ?? "") !== "";
+  if (sameClientType(value, fixed)) return false;
+  return stored === void 0 || !sameClientType(value, stored);
+}
+var MODELSCOPE_PROVIDER_ID = "modelscope";
+var MODEL_PROVIDERS = [
+  {
+    id: "tokendance",
+    label: "TokenDance",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://tokendance.space/keys",
+    modelsUrl: "https://tokendance.space/models",
+    gatewayBaseUrl: TOKENDANCE_BASE_URL,
+    // Every preset speaks Chat Completions, and the group takes models added by hand
+    // (`addable`), which follow the group: a new Project's `[providers.tokendance]` carries
+    // this pin, so such a row resolves to it.
+    clientType: "openai-chat",
+    addable: true,
+    recommended: true,
+    // https://tokendance.space/docs/api-key-oauth
+    oauth: {
+      authorizeUrl: "https://tokendance.space/auth",
+      exchangeUrl: "https://tokendance.space/portal/api/v1/auth/keys",
+      keyName: "PenguinHarness"
+    },
+    // https://tokendance.space/docs/open-api (the account's wallet, in micro-yuan)
+    balance: {
+      url: "https://tokendance.space/portal/api/v1/user/balance",
+      format: "tokendance"
+    }
+  },
+  {
+    id: PENGUIN_GO_PROVIDER_ID,
+    label: "Penguin Go",
+    // The authorization flow writes the relay key once, on the group's `[providers.penguin-go]`
+    // table, and that key is the only one the group's rows use: a third-party adapter takes
+    // keys from the config alone, and every row names the relay, so none is handed the
+    // environment's. The env pair records what the mmsp client reads, as a fact. Every row names the relay and speaks MMSP, so a new
+    // Project writes both on the group; the protocol is fixed (fixedClientType).
+    envKey: "MMSP_API_KEY",
+    envBaseUrlKey: "MMSP_BASE_URL",
+    apiKeyUrl: "https://go.penguin.ooo/",
+    modelsUrl: "https://go.penguin.ooo/",
+    bridgeAuth: { flow: "penguin-go" },
+    balance: {
+      url: "https://go.penguin.ooo/api/me/by-key",
+      format: "penguin-go"
+    }
+  },
+  {
+    // One key serves the whole group, but its models sit on four protocols (see the group's
+    // block comment in MODEL_CATALOG), so no group-level pin: a new Project writes each row's
+    // protocol on the row, and the Messages or Google base on the rows that use it. The env pair
+    // records what its Chat Completions majority's client reads. Like every gateway's, its rows
+    // carry a base URL, so a keyless row is refused: the key goes on the group or the row.
+    id: "opencode-go",
+    label: "OpenCode Go",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://opencode.ai/v2/docs/console/api/inference/#authentication",
+    modelsUrl: "https://opencode.ai/v2/docs/console/go/",
+    gatewayBaseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    id: "deepseek",
+    label: "DeepSeek",
+    envKey: "DEEPSEEK_API_KEY",
+    envBaseUrlKey: "DEEPSEEK_BASE_URL",
+    apiKeyUrl: "https://platform.deepseek.com/api_keys",
+    modelsUrl: "https://api-docs.deepseek.com/quick_start/pricing",
+    // https://api-docs.deepseek.com/api/get-user-balance
+    balance: { url: `${DEEPSEEK_BASE_URL}/user/balance`, format: "deepseek" }
+  },
+  {
+    id: "openrouter",
+    label: "OpenRouter",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://openrouter.ai/workspaces/default/keys",
+    modelsUrl: "https://openrouter.ai/models",
+    gatewayBaseUrl: OPENROUTER_BASE_URL,
+    // The whole group speaks the Responses API, which OpenRouter serves at the preset base
+    // URL below for every upstream — presets and user-added entries alike.
+    clientType: "openai-responses",
+    // The catalog lists a fraction of what OpenRouter serves, so the group takes more.
+    addable: true
+  },
+  {
+    id: "fireworks",
+    label: "Fireworks AI",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://app.fireworks.ai/settings/users/api-keys",
+    modelsUrl: "https://app.fireworks.ai/models",
+    gatewayBaseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    id: "google",
+    label: "Google Gemini",
+    envKey: "GEMINI_API_KEY",
+    envBaseUrlKey: "GEMINI_BASE_URL",
+    apiKeyUrl: "https://aistudio.google.com/api-keys",
+    modelsUrl: "https://ai.google.dev/gemini-api/docs/models"
+  },
+  {
+    id: "openai",
+    label: "OpenAI",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://platform.openai.com/api-keys",
+    modelsUrl: "https://platform.openai.com/docs/models"
+  },
+  {
+    id: "anthropic",
+    label: "Anthropic",
+    envKey: "ANTHROPIC_API_KEY",
+    envBaseUrlKey: "ANTHROPIC_BASE_URL",
+    apiKeyUrl: "https://platform.claude.com/settings/keys",
+    modelsUrl: "https://docs.claude.com/en/docs/about-claude/models/overview"
+  },
+  {
+    id: "siliconflow",
+    label: "SiliconFlow",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://cloud.siliconflow.cn/me/account/ak",
+    modelsUrl: "https://cloud.siliconflow.cn/models",
+    gatewayBaseUrl: SILICONFLOW_BASE_URL,
+    // As TokenDance: Chat Completions throughout, and models added by hand follow the group.
+    clientType: "openai-chat",
+    addable: true
+  },
+  {
+    id: "atlascloud",
+    label: "Atlas Cloud",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://atlascloud.ai/docs/api-keys",
+    modelsUrl: "https://atlascloud.ai/models",
+    gatewayBaseUrl: ATLASCLOUD_BASE_URL,
+    // As SiliconFlow: Chat Completions throughout, and the catalog lists a fraction of what
+    // the gateway serves, so the group takes models added by hand, which follow the group.
+    clientType: "openai-chat",
+    addable: true
+  },
+  {
+    id: "zhipu",
+    label: "Z.AI (GLM)",
+    envKey: "ZAI_API_KEY",
+    envBaseUrlKey: "ZAI_BASE_URL",
+    apiKeyUrl: "https://open.bigmodel.cn/apikey/platform",
+    modelsUrl: "https://docs.z.ai/guides/overview/pricing"
+  },
+  {
+    id: "moonshot",
+    label: "Moonshot (Kimi)",
+    envKey: "MOONSHOT_API_KEY",
+    envBaseUrlKey: "MOONSHOT_BASE_URL",
+    apiKeyUrl: "https://platform.kimi.com/console/api-keys",
+    modelsUrl: "https://platform.kimi.com/docs/pricing"
+  },
+  {
+    id: "minimax",
+    label: "MiniMax",
+    envKey: "MINIMAX_API_KEY",
+    envBaseUrlKey: "MINIMAX_BASE_URL",
+    // The pay-as-you-go key page; a Token Plan Subscription Key (Billing > Token Plan) works
+    // against the same endpoint, so the group is not tied to either billing mode.
+    apiKeyUrl: "https://platform.minimax.io/user-center/basic-information/interface-key",
+    modelsUrl: "https://platform.minimax.io/docs/guides/models-intro"
+  },
+  {
+    id: "qwen-pay-as-you-go",
+    label: "Qwen Pay-As-You-Go",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://platform.qianwenai.com/docs/api-reference/preparation/api-key",
+    modelsUrl: "https://www.qianwenai.com/models",
+    gatewayBaseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    id: "qwen-token-plan",
+    label: "Qwen Token Plan",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://platform.qianwenai.com/pricing/token-plan",
+    modelsUrl: "https://platform.qianwenai.com/docs/token-plan/personal/token-plan-personal-overview",
+    gatewayBaseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  {
+    // ModelScope's own inference service: OpenAI-compatible, reselling other vendors' models
+    // under `org/Model-Name` ids. Two things set it apart from the gateways above.
+    //
+    // Its credential is the account's access token rather than a minted API key, and the
+    // group obtains one through the authorization bridge instead of sending the user to copy
+    // it out of the console (bridgeAuth below; the console link stays for anyone who already
+    // has a token). ModelScope is the vendor here as well as the gateway — the bridge exists
+    // because its OAuth requires a client secret the desktop App cannot keep locally, not
+    // because inference requests are routed through that bridge.
+    //
+    // Every preset row names the endpoint and one protocol in the catalog, so a new Project
+    // writes both on the group's `[providers.modelscope]` table and the rows follow it.
+    id: MODELSCOPE_PROVIDER_ID,
+    label: "ModelScope",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    apiKeyUrl: "https://www.modelscope.cn/my/myaccesstoken",
+    modelsUrl: "https://www.modelscope.cn/models",
+    gatewayBaseUrl: MODELSCOPE_BASE_URL,
+    bridgeAuth: { flow: "modelscope" }
+  },
+  {
+    // Self-hosted: the user runs the server, so there is no console to mint a key at
+    // (apiKeyUrl) and no endpoint every user shares (gatewayBaseUrl). The group's rows name the
+    // address a server started on this machine answers on, which a new Project stores as the
+    // group's base URL. modelsUrl points at the recipe index, which is where the served ids in
+    // this group are documented.
+    id: "vllm",
+    label: "vLLM",
+    // vLLM's own pair, which MMSP's vLLM client reads, and never OpenAI's.
+    envKey: "VLLM_API_KEY",
+    envBaseUrlKey: "VLLM_BASE_URL",
+    modelsUrl: "https://recipes.vllm.ai/",
+    clientType: VLLM_CLIENT_TYPE,
+    // The served ids are the user's own, so this group keeps taking models added by hand.
+    addable: true
+  },
+  {
+    id: "custom",
+    label: "Custom",
+    envKey: "OPENAI_API_KEY",
+    envBaseUrlKey: "OPENAI_BASE_URL",
+    addable: true
+  }
+];
+function cny(cacheRead, cacheWrite, output) {
+  const r = (v) => Math.round(v / 7 * 1e6) / 1e6;
+  return usd(r(cacheRead), r(cacheWrite), r(output));
+}
+function usd(cacheRead, cacheWrite, output) {
+  return { unit: "usd_per_mtok", cache_read: cacheRead, cache_write: cacheWrite, output };
+}
+function offPeakAt(schedule, now) {
+  const local = new Date(now.getTime() + schedule.utcOffsetMinutes * 6e4);
+  const isoDay = local.getUTCDay() === 0 ? 7 : local.getUTCDay();
+  if (!schedule.peakDays.includes(isoDay)) return true;
+  const hour = local.getUTCHours() + local.getUTCMinutes() / 60;
+  return !schedule.peakHours.some(([from, to]) => hour >= from && hour < to);
+}
+var DEEPSEEK_OFF_PEAK = {
+  rate: 0.5,
+  utcOffsetMinutes: 480,
+  peakDays: [1, 2, 3, 4, 5],
+  peakHours: [
+    [9, 12],
+    [14, 18]
+  ]
+};
+var QWEN_OFF_PEAK = {
+  rate: 0.5,
+  utcOffsetMinutes: 480,
+  peakDays: [1, 2, 3, 4, 5, 6, 7],
+  peakHours: [[8, 22]]
+};
+function offPeakScheduledRefs() {
+  const bySchedule = /* @__PURE__ */ new Map();
+  for (const entry of MODEL_CATALOG) {
+    const schedule = entry.offPeakDiscount;
+    if (schedule === void 0) continue;
+    const group = bySchedule.get(schedule) ?? { schedule, refs: [] };
+    group.refs.push({ provider: entry.provider, modelId: entry.modelId });
+    bySchedule.set(schedule, group);
+  }
+  return [...bySchedule.values()];
+}
+function effectivePricing(entry, now = /* @__PURE__ */ new Date()) {
+  const { pricing } = entry;
+  if (pricing === void 0) return pricing;
+  const tier = entry.offPeakDiscount !== void 0 && offPeakAt(entry.offPeakDiscount, now) ? entry.offPeakDiscount.rate : 0;
+  const kept = (1 - tier) * (1 - (entry.discount ?? 0));
+  if (kept === 1) return pricing;
+  const off = (v) => Math.round(v * kept * 1e6) / 1e6;
+  return {
+    unit: pricing.unit,
+    cache_read: off(pricing.cache_read),
+    cache_write: off(pricing.cache_write),
+    output: off(pricing.output)
+  };
+}
+var MODEL_CATALOG = [
+  // -- DeepSeek (official CNY pricing: cache hit / cache miss / output). Prices and model
+  // names re-read 2026-09-16 from api-docs.deepseek.com/zh-cn/quick_start/pricing, which now
+  // lists exactly two names, both with a 1M context window and a 384K output cap:
+  // `deepseek-flash` (DeepSeek-V4.1-Flash, image input) on the peak tier CNY 0.04 / 2 / 8, and
+  // `deepseek-v4-pro` (DeepSeek-V4-Pro-0813, text only) at 0.30 / 9 / 27. The rows store the
+  // PEAK tier and declare DEEPSEEK_OFF_PEAK, which halves every bucket outside Beijing weekday
+  // 09:00-12:00 and 14:00-18:00 — so both tiers are billed at the rate actually in force,
+  // rather than one of them being approximated by the other.
+  // The retired V4 Flash names, `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp`, are
+  // still accepted and served by V4.1 Flash at the Flash price and on the same schedule, but
+  // the page no longer lists them: they stay below as retired rows, which are not presets. --
+  {
+    // DeepSeek's pricing page names this model `deepseek-flash` (model version
+    // DeepSeek-V4.1-Flash, released as of the 2026-09-10 read): 1M context, image input, and
+    // the Flash series peak price and off-peak schedule.
+    modelId: "deepseek-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "deepseek",
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    wasDefault: true
+  },
+  {
+    // Retired 2026-09-16 (see ModelCatalogEntry.retired): the default model of new Projects
+    // from 0.2.0 to 0.2.8. DeepSeek still accepts the id and serves it from V4.1 Flash at the
+    // Flash price, off-peak schedule included. Text only: MMSP's DeepSeek client refuses
+    // image parts for this id.
+    modelId: "deepseek-v4-flash",
+    displayName: "DeepSeek V4 Flash",
+    provider: "deepseek",
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    retired: true,
+    wasDefault: true
+  },
+  {
+    // Retired 2026-09-16 (see ModelCatalogEntry.retired): the default model of new Projects in
+    // 0.2.9, served from V4.1 Flash at the Flash price like deepseek-v4-flash above.
+    modelId: "deepseek-v4-flash-vision-exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "deepseek",
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    retired: true,
+    wasDefault: true
+  },
+  {
+    // DeepSeek's pricing page names the model behind this id DeepSeek-V4-Pro-0813 (read
+    // 2026-09-16), which is what the display name says — the same 0813 release the gateway
+    // groups sell under dated ids. Text only. DeepSeek states that V4 Pro stays available
+    // after 2026-09-14 with its billing unchanged, so the row keeps the V4 Pro peak price on
+    // the shared off-peak schedule. MMSP's DeepSeek client also matches this bare id against
+    // its text-only deny-list /^deepseek-v4-(flash|pro)(-\d{4})?$/, so image parts never
+    // leave the harness for it.
+    modelId: "deepseek-v4-pro",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: "deepseek",
+    contextWindow: 1e6,
+    pricing: cny(0.3, 9, 27),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    wasDefault: true
+  },
+  // -- OpenRouter (gateway: OpenAI-compatible protocol, preset base URL). Prices re-read in
+  // one pass on 2026-08-07 from the models API (/api/v1/models; the API is authoritative
+  // where a model's web page disagrees); the DeepSeek rows, the rows added on 2026-08-18
+  // (grok-4.6, deepseek-v4-pro-0813, glm-5.3, the openai/* additions) and every
+  // pre-existing openai/* and google/* row were re-read on 2026-08-18 from the model
+  // pages and the per-model endpoints API. The rows added on 2026-10-10 — the four Claude 5.x
+  // rows, openai/gpt-6.1-sol, stepfun/step-5-preview, x-ai/grok-4.7,
+  // xiaomi/mimo-v2.6-flash and xiaomi/mimo-v2.6-pro — were read that day from the models API and the per-model
+  // endpoints API, whose default endpoint runs no promotion (`discount: 0`) on any of them;
+  // their context windows are the listing's `top_provider.context_length`, as on the
+  // mimo-v2.5 and step-3.7-flash rows.
+  //
+  // Protocol: every row pins `openai-responses`. OpenRouter serves the Responses API at
+  // {base}/responses for every upstream, at the same https://openrouter.ai/api/v1 base URL
+  // the rows already carry, and the group pins the same protocol, which a new Project writes on
+  // `[providers.openrouter]`, so an entry added to it by hand follows it. The generic Responses client sends a text-only tool result as a plain
+  // string and replays reasoning items only where the upstream returned them.
+  //
+  // Price buckets: cache_read stores the published input_cache_read
+  // (falling back to the input price for the rows without one — the :free rows and the
+  // GPT Pro tiers, which publish no cache discount); cache_write stores
+  // input_cache_write only when it is a genuine per-token write premium (the Anthropic, GPT
+  // and qwen3.8-max rows, 1.25x input) —
+  // Gemini's field is an hourly cache-STORAGE rate, not a per-token price, so those rows
+  // keep the input price — and otherwise also carries the input price. The :free tier and
+  // the openrouter/free Free Models Router store a genuine $0 price (not "unknown"), so
+  // costs correctly compute to 0. GPT models are uniformly vision-capable (OpenAI
+  // product-line policy) even where the gateway page omits the modality.
+  //
+  // Discounts: what these rows record is what OpenRouter actually BILLS, so a gateway
+  // promotion is stored at its discounted rate (unlike the direct-vendor rows, which keep the
+  // list price). The Gemini 3.x Flash rows are the exception: the promotion they sit on is
+  // Google's own dated launch discount rather than the gateway's, so they keep the list price
+  // and declare it in `discount`, exactly as their direct-vendor twins do. The endpoints API
+  // exposes the running promotion as `pricing.discount` on the default endpoint; rows sitting
+  // on one say so and name the rate to restore, because a lapsed promotion silently doubles
+  // the real cost — that is exactly how the gpt-5.6-terra and gpt-5.6-luna rows drifted 2x
+  // low before the 2026-08-18 re-read. --
+  {
+    // $10 input / $12.50 cache write / $0.25 cache read / $50 output, the cache read at 2.5% of
+    // input as on Anthropic's own list; 1M window, image input.
+    modelId: "anthropic/claude-fable-5.1",
+    displayName: "Claude Fable 5.1",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.25, 12.5, 50),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "anthropic/claude-fable-5",
+    displayName: "Claude Fable 5",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(1, 12.5, 50),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // $0.10 input / $0.125 cache write / $0.01 cache read / $0.50 output for prompts up to
+    // 100,000 tokens; the listing's `overrides` bill every rate 5x from 100,000 prompt tokens,
+    // Anthropic's own tiering. The base tier is recorded, as for the other tiered rows.
+    modelId: "anthropic/claude-haiku-5.5",
+    displayName: "Claude Haiku 5.5",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.01, 0.125, 0.5),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // $4 input / $5 cache write / $0.20 cache read / $20 output.
+    modelId: "anthropic/claude-opus-5.5",
+    displayName: "Claude Opus 5.5",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.2, 5, 20),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "anthropic/claude-opus-5",
+    displayName: "Claude Opus 5",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 6.25, 25),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "anthropic/claude-opus-4.8",
+    displayName: "Claude Opus 4.8",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 6.25, 25),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "anthropic/claude-opus-4.7",
+    displayName: "Claude Opus 4.7",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 6.25, 25),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // $2 input / $2.50 cache write / $0.10 cache read / $10 output. The cache read is half
+    // Sonnet 5's $0.20: Anthropic prices it at 5% of input on this generation (MMSP 0.5.2
+    // moved its registry row from $0.20), and OpenRouter bills the same.
+    modelId: "anthropic/claude-sonnet-5.5",
+    displayName: "Claude Sonnet 5.5",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.1, 2.5, 10),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "anthropic/claude-sonnet-5",
+    displayName: "Claude Sonnet 5",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.2, 2.5, 10),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Listed 2026-09-10 from the models API: the base price is the PEAK tier and
+    // `pricing.overrides` bill exactly half on weekends and outside weekday UTC 01:00-04:00 /
+    // 06:00-10:00 — DeepSeek's own Beijing windows — so the row stores the peak price and
+    // declares DEEPSEEK_OFF_PEAK, exactly as the direct row does. Context window and image
+    // input from the same listing.
+    modelId: "deepseek/deepseek-v4.1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(6e-3, 0.3, 1.2),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The 1017 snapshot, served by DeepSeek's own endpoint (MMSP 0.5.5's registry, models API
+    // read 2026-10-19): $0.15 / $0.60 / $0.003 off-peak, double in DeepSeek's weekday windows,
+    // so the row stores that peak and declares DEEPSEEK_OFF_PEAK like the row above.
+    modelId: "deepseek/deepseek-v4.1-flash-1017",
+    displayName: "DeepSeek V4.1 Flash 1017",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(6e-3, 0.3, 1.2),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "deepseek/deepseek-v4-flash-0731",
+    displayName: "DeepSeek V4 Flash 0731",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.0157192, 0.078596, 0.157192),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "deepseek/deepseek-v4-flash",
+    displayName: "DeepSeek V4 Flash",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.0168, 0.0679, 0.168),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // DeepSeek serves this one alone on OpenRouter, so the stored rates are its own published
+    // USD list ($0.22 / $0.66 / $0.007 cache read), and the context window is that endpoint's
+    // 1,048,576. The row keeps that flat price as the API lists it: as of the 2026-09-10 read
+    // the listing carries no `pricing.overrides`, so unlike the deepseek-v4.1-flash row above
+    // there is no peak/off-peak split to record here.
+    modelId: "deepseek/deepseek-v4-flash-vision-exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(7e-3, 0.22, 0.66),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The 0813 general-availability release of DeepSeek V4 Pro (OpenRouter listing dated
+    // 2026-08-12); the default routed endpoint is DeepSeek's own API, so the price matches
+    // the official USD list, and the context window is the default endpoint's 1,048,576.
+    modelId: "deepseek/deepseek-v4-pro-0813",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.022, 0.66, 1.98),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "deepseek/deepseek-v4-pro",
+    displayName: "DeepSeek V4 Pro",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.022, 0.66, 1.98),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Standard tier from openrouter.ai/google/gemini-4-argon (2026-10-14): Google's
+    // $4 input / $20 output / $0.20 cache hit, 50% off through 2026-11-30. Argon supports
+    // implicit caching only, so cache_write records input rather than cache storage.
+    modelId: "google/gemini-4-argon",
+    displayName: "Gemini 4 Argon",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.2, 4, 20),
+    discount: 0.5,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Same Gemini cache conventions as the gemini-3.6-flash row below. OpenRouter passes
+    // Google's launch discount through, so it bills $0.075/$0.75/$3.75 — the list price less
+    // 50%, declared in `discount` so the $0.15/$1.50/$7.50 list survives the promotion (same
+    // treatment as the 3.6 row below). Gemini 3.7 Flash is no longer listed beside it: Google
+    // retires that id and reroutes it to 3.8 Flash.
+    modelId: "google/gemini-3.8-flash",
+    displayName: "Gemini 3.8 Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 7.5),
+    discount: 0.5,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // cache_read is billed as its own bucket in the cost center, and an input-priced
+    // cache_read would overstate cache-heavy Gemini spend 10x; cache_write repeats the input
+    // price (see the block comment — Gemini publishes storage-per-hour, not per-token write),
+    // matching the direct-vendor Gemini rows below. OpenRouter bills $0.075/$0.75/$3.75
+    // today (endpoints API, read 2026-09-09): it reports that halved rate as its plain price
+    // with `discount: 0`, but it is Google's launch discount passed through and ends with it
+    // on 2026-12-31, so this row records the list price and the promotion the way its
+    // siblings do.
+    modelId: "google/gemini-3.6-flash",
+    displayName: "Gemini 3.6 Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 7.5),
+    discount: 0.5,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The 3.5 tier's own list price — $1.50 input / $9.00 output / $0.15 cache hit on Google's
+    // page, with no launch discount — and OpenRouter's default endpoint bills exactly that
+    // (`discount: 0`, read 2026-09-09). The $9 output is the tier, not a promotion to declare.
+    modelId: "google/gemini-3.5-flash",
+    displayName: "Gemini 3.5 Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 9),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Same published-cache-price convention as gemini-3.6-flash above (2026-07-22: $0.03/mtok
+    // cache hit, $0.30 input, $2.50 output; re-read 2026-09-09, unchanged and `discount: 0`).
+    modelId: "google/gemini-3.5-flash-lite",
+    displayName: "Gemini 3.5 Flash-Lite",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.03, 0.3, 2.5),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The paid listing, which replaced the free variant when its free period ended (models API
+    // read 2026-10-23): $0.15 input / $0.60 output / $0.015 cache read, 1,048,576 tokens, image
+    // input, MiniMax's own list price.
+    modelId: "minimax/minimax-m3.1-flash",
+    displayName: "MiniMax M3.1 Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.015, 0.15, 0.6),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // No official separate cache price published: cache_read uses the standard input price (no discount assumed).
+    modelId: "minimax/minimax-m3",
+    displayName: "MiniMax M3",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.06, 0.3, 1.2),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Read 2026-10-24 from the models API and the per-model endpoints API: $2 input / $10
+    // output / $0.20 cache read over 1,048,576 tokens, image input, served by Moonshot AI alone
+    // and with no promotion on the default endpoint (`discount: 0`). The listing's
+    // `moonshotai/highspeed` endpoint ($4 / $20 / $0.20) is a provider route under this same id,
+    // not a model of its own.
+    modelId: "moonshotai/kimi-k3.1-pro",
+    displayName: "Kimi K3.1 Pro",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.2, 2, 10),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "moonshotai/kimi-k3",
+    displayName: "Kimi K3",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.3, 3, 15),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "moonshotai/kimi-k2.6",
+    displayName: "Kimi K2.6",
+    provider: "openrouter",
+    contextWindow: 262144,
+    pricing: usd(0.0992, 0.589, 2.48),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "nvidia/nemotron-3-ultra-550b-a55b:free",
+    displayName: "Nemotron 3 Ultra (free)",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  // The openai/* rows below mirror the direct OpenAI group one-for-one. Their context
+  // windows are OpenRouter's published 1,050,000 / 400,000, matching the direct rows.
+  {
+    // Read 2026-10-10: the default endpoint is OpenAI's own, at $2 input / $0.10 cached input
+    // / $2.50 cache write / $10 output with `discount: 0`, OpenAI's list. The listing's
+    // `overrides` from 272,000 prompt tokens (2x prompt and cache, 1.5x completion) are the
+    // long-context tier, which is not recorded, as on the direct row.
+    modelId: "openai/gpt-6.1-sol",
+    displayName: "GPT-6.1 Sol",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.1, 2.5, 10),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Read 2026-09-09 from the models API and the per-model endpoints API: the default
+    // endpoint is OpenAI's own, listed at $10 input / $1 cached input / $12.5 cache write /
+    // $50 output with `discount: 0`, so the list price is what OpenRouter bills. The
+    // endpoints API also publishes `overrides` above 272,000 prompt tokens (2x prompt and
+    // cache, 1.5x completion); as on the direct row, only the base tier is recorded.
+    modelId: "openai/gpt-6-astra",
+    displayName: "GPT-6 Astra",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(1, 12.5, 50),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // https://developers.openai.com/api/docs/models/gpt-6-luna (2026-10-14).
+    // Base tier: above 272K input, input/cache rates double and output is 1.5x.
+    modelId: "openai/gpt-6-luna",
+    displayName: "GPT-6 Luna",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.01, 0.125, 0.5),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The 50% promotion this row used to store has ended: OpenRouter now bills the full
+    // $0.20/$1.20 rate (endpoints API `discount: 0`), so the stored rates doubled on the
+    // 2026-08-18 re-read.
+    modelId: "openai/gpt-5.6-luna",
+    displayName: "GPT-5.6 Luna",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.02, 0.25, 1.2),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Running a `discount: 0.5` promotion as of 2026-08-18, so OpenRouter bills half the
+    // $0.50/$6.25/$30 list price — restore the list rates when the promotion ends.
+    modelId: "openai/gpt-5.6-sol",
+    displayName: "GPT-5.6 Sol",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.25, 3.125, 15),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Same lapsed promotion as the luna row: now billed at the full $2/$12 rate.
+    modelId: "openai/gpt-5.6-terra",
+    displayName: "GPT-5.6 Terra",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.2, 2.5, 12),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "openai/gpt-5.5",
+    displayName: "GPT-5.5",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.5, 5, 30),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // No published cache discount (the Pro tiers bill cached input at the standard rate), so
+    // cache_read carries the input price — same convention as the direct gpt-5.5-pro row.
+    modelId: "openai/gpt-5.5-pro",
+    displayName: "GPT-5.5 Pro",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(30, 30, 180),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "openai/gpt-5.4",
+    displayName: "GPT-5.4",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(0.25, 2.5, 15),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "openai/gpt-5.4-mini",
+    displayName: "GPT-5.4 mini",
+    provider: "openrouter",
+    contextWindow: 4e5,
+    pricing: usd(0.075, 0.75, 4.5),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "openai/gpt-5.4-nano",
+    displayName: "GPT-5.4 nano",
+    provider: "openrouter",
+    contextWindow: 4e5,
+    pricing: usd(0.02, 0.2, 1.25),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // No published cache discount, as with gpt-5.5-pro above.
+    modelId: "openai/gpt-5.4-pro",
+    displayName: "GPT-5.4 Pro",
+    provider: "openrouter",
+    contextWindow: 105e4,
+    pricing: usd(30, 30, 180),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // OpenRouter's unified Free Models Router: each request is routed to a random free model
+    // currently on OpenRouter, filtered by the features the request needs (tool calling,
+    // structured outputs, ...). Routed targets vary, so the context window is a deliberately
+    // conservative figure rather than any single target's real window: it keeps the 75%
+    // compaction clamp meaningful (compaction fires at 96000) and reduces hard context-length
+    // 400s on small-window targets. supportsVision stays false deliberately: the harness must
+    // not send images to a router whose target may be text-only.
+    modelId: "openrouter/free",
+    displayName: "Free Models Router",
+    provider: "openrouter",
+    contextWindow: 128e3,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Read 2026-09-16 from the models API: $0.214 input / $2.55 output with a published $0.15
+    // input_cache_read, a 1,000,000-token context window, and text, image and video input.
+    modelId: "qwen/qwen3.8-27b",
+    displayName: "Qwen 3.8 27B",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.15, 0.214, 2.55),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "qwen/qwen3.8-max",
+    displayName: "Qwen 3.8 Max",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.25, 2.5, 6),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "qwen/qwen3.6-35b-a3b",
+    displayName: "Qwen 3.6 35B A3B",
+    provider: "openrouter",
+    contextWindow: 262144,
+    pricing: usd(0.05, 0.14, 1),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Recursive's Flint, listed 2026-10-21 (models API read 2026-10-23): free, 262,144 tokens,
+    // up to 235,929 output tokens, text input only, with reasoning and tool calls.
+    modelId: "recursive/flint-v1.0:free",
+    displayName: "Flint (free)",
+    provider: "openrouter",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // StepFun's own endpoint is the only one: $1 input / $2.70 output with a published $0.05
+    // input_cache_read and no write premium, so cache_write carries the input price. 1,000,000
+    // window; text, image and video input.
+    modelId: "stepfun/step-5-preview",
+    displayName: "Step 5 Preview",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.05, 1, 2.7),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // No official separate cache price published: cache_read uses the standard input price.
+    modelId: "stepfun/step-3.7-flash",
+    displayName: "Step 3.7 Flash",
+    provider: "openrouter",
+    contextWindow: 256e3,
+    pricing: usd(0.04, 0.2, 1.15),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Tencent's Hy4 preview, read 2026-08-28 from its OpenRouter page and the models API:
+    // $0.834 input / $2.501 output with a published $0.042 input_cache_read, a 1,048,576
+    // context window, and text-only modalities in and out. The same upstream model is sold
+    // in the TokenDance group at that gateway's own CNY rate; the two rows are priced by
+    // their sellers and are not copies of one another.
+    modelId: "tencent/hy4-preview",
+    displayName: "Hy4 preview",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.042, 0.834, 2.501),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "tencent/hy3",
+    displayName: "Hy3",
+    provider: "openrouter",
+    contextWindow: 262144,
+    pricing: usd(0.033, 0.132, 0.528),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Thinking Machines Lab's Inkling (released 2026-07-14): multimodal (image + audio
+    // input). Specs from its OpenRouter page; pricing from the models API (2026-08-07),
+    // which publishes $1 input (the page shows $0.95) and a $0.17 cached-input price.
+    modelId: "thinkingmachines/inkling",
+    displayName: "Inkling",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.17, 1, 4.05),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // Grok 4.7 keeps Grok 4.6's $2 input / $6 output / $0.50 cache hit and its 500,000 window,
+    // with image input. The listing's `overrides` double every rate from 200,000 prompt tokens;
+    // the base tier is recorded.
+    modelId: "x-ai/grok-4.7",
+    displayName: "Grok 4.7",
+    provider: "openrouter",
+    contextWindow: 5e5,
+    pricing: usd(0.5, 2, 6),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // xAI's Grok 4.6 (OpenRouter listing dated 2026-08-12): same $2/$6 input/output rates as
+    // Grok 4.5 with a raised $0.50 cache-hit price.
+    modelId: "x-ai/grok-4.6",
+    displayName: "Grok 4.6",
+    provider: "openrouter",
+    contextWindow: 5e5,
+    pricing: usd(0.5, 2, 6),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "x-ai/grok-4.5",
+    displayName: "Grok 4.5",
+    provider: "openrouter",
+    contextWindow: 5e5,
+    pricing: usd(0.3, 2, 6),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // $0.14 input / $0.28 output with a published $0.0028 input_cache_read, the same rates as
+    // MiMo-V2.5; text, image, video and audio input.
+    modelId: "xiaomi/mimo-v2.6-flash",
+    displayName: "MiMo-V2.6-Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(28e-4, 0.14, 0.28),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // $0.435 input / $0.87 output with a published $0.0036 input_cache_read (models API and
+    // the per-model endpoints API, read 2026-10-10: Xiaomi's own endpoint, no discount — the
+    // USD face of TokenDance's CNY 0.025 / 3 / 6); text, image, video and audio input.
+    modelId: "xiaomi/mimo-v2.6-pro",
+    displayName: "MiMo-V2.6-Pro",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(36e-4, 0.435, 0.87),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "xiaomi/mimo-v2.5",
+    displayName: "MiMo-V2.5",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(28e-4, 0.14, 0.28),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The gateway listing of the direct glm-5.3 row below; OpenRouter's single Z.AI endpoint
+    // passes Z.AI's published price straight through (no discount), which is why the two
+    // rows agree to the cent. Text-only, per the listing's modalities.
+    modelId: "z-ai/glm-5.3",
+    displayName: "GLM-5.3",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    // The gateway listing of the direct glm-5.3-flash row below. Stored at the rate OpenRouter's
+    // default endpoint bills: DeepInfra (fp4), running `discount: 0.5` on the 0.03 / 0.15 / 0.5
+    // list (endpoints API, re-read 2026-09-16) — restore the list rates when that endpoint's
+    // promotion ends. Z.AI's own 50% promotion is a different one: it ended at 2026-09-09
+    // 16:00 UTC, and the Z.AI endpoint on OpenRouter bills the list price since. The listing
+    // takes text, images and video, and the generic Responses client it pins carries image
+    // parts through.
+    modelId: "z-ai/glm-5.3-flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.015, 0.075, 0.25),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  {
+    modelId: "z-ai/glm-5.2",
+    displayName: "GLM-5.2",
+    provider: "openrouter",
+    contextWindow: 1e6,
+    pricing: usd(0.1261, 0.679, 2.134),
+    supportsVision: false,
+    clientType: "openai-responses",
+    baseUrl: OPENROUTER_BASE_URL
+  },
+  // -- Fireworks AI (gateway, standard serverless USD pricing: cached input / uncached
+  // input / output from each model's page; API ids use the accounts/fireworks/models/<slug>
+  // form, where Fireworks spells a version's dot as `p`: deepseek-v4p1-flash, glm-5p3). The
+  // DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813, GLM-5.3, GLM-5.3 Flash and Qwen 3.8 Max rows
+  // were added 2026-09-16 from their fireworks.ai/models pages. Fireworks publishes no context
+  // length for qwen3p8-max, so that row records the vendor's 1M window, as the other
+  // qwen3.8-max rows do. --
+  {
+    // Fireworks' listing of the model DeepSeek serves directly as `deepseek-flash`: 1,048,576
+    // context, image input, and the single serverless price its page lists, so no schedule.
+    modelId: "accounts/fireworks/models/deepseek-v4p1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "fireworks",
+    contextWindow: 1048576,
+    pricing: usd(7e-3, 0.22, 0.66),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/deepseek-v4-flash-0731",
+    displayName: "DeepSeek V4 Flash 0731",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.028, 0.14, 0.28),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/deepseek-v4-flash",
+    displayName: "DeepSeek V4 Flash",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.03, 0.14, 0.28),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/deepseek-v4-pro-0813",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: "fireworks",
+    contextWindow: 1048576,
+    pricing: usd(0.044, 1.32, 3.96),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/deepseek-v4-pro",
+    displayName: "DeepSeek V4 Pro",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.15, 1.74, 3.48),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/glm-5p3",
+    displayName: "GLM-5.3",
+    provider: "fireworks",
+    contextWindow: 1048576,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    // Natively multimodal like every other GLM-5.3 Flash row, and this group's generic
+    // openai-chat client forwards image_url parts, so image input works on this path.
+    modelId: "accounts/fireworks/models/glm-5p3-flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "fireworks",
+    contextWindow: 1048576,
+    pricing: usd(0.03, 0.15, 0.5),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/glm-5p2",
+    displayName: "GLM-5.2",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.14, 1.4, 4.4),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    // Thinking Machines Lab's Inkling (released 2026-07-14): multimodal (image + audio
+    // input); specs and serverless pricing from its Fireworks model page (2026-08-06).
+    modelId: "accounts/fireworks/models/inkling",
+    displayName: "Inkling",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.17, 1, 4.05),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/kimi-k3",
+    displayName: "Kimi K3",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.3, 3, 15),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/kimi-k2p7-code",
+    displayName: "Kimi K2.7 Code",
+    provider: "fireworks",
+    contextWindow: 262144,
+    pricing: usd(0.19, 0.95, 4),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/minimax-m3",
+    displayName: "MiniMax M3",
+    provider: "fireworks",
+    contextWindow: 524288,
+    pricing: usd(0.06, 0.3, 1.2),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  {
+    modelId: "accounts/fireworks/models/qwen3p8-max",
+    displayName: "Qwen 3.8 Max",
+    provider: "fireworks",
+    contextWindow: 1e6,
+    pricing: usd(0.25, 2, 6),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: FIREWORKS_BASE_URL
+  },
+  // -- SiliconFlow (gateway, official CNY pricing: cache hit / input / output) --
+  {
+    // CNY 1 input / 4 output / 0.02 cache hit over 1,048,576 tokens, image input (MMSP 0.5.5's
+    // registry, SiliconFlow's price list).
+    modelId: "deepseek-ai/DeepSeek-V4.1-Flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(0.02, 1, 4),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    // SiliconFlow repriced it to CNY 1.5 / 4.5 / 0.15 (MMSP 0.5.5's registry).
+    modelId: "deepseek-ai/DeepSeek-V4-Flash",
+    displayName: "DeepSeek V4 Flash",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(0.15, 1.5, 4.5),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    modelId: "deepseek-ai/DeepSeek-V4-Pro",
+    displayName: "DeepSeek V4 Pro",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(1, 12, 24),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    modelId: "meituan-longcat/LongCat-2.0",
+    displayName: "LongCat 2.0",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(0.1, 5, 20),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    modelId: "moonshotai/Kimi-K2.7-Code",
+    displayName: "Kimi K2.7 Code",
+    provider: "siliconflow",
+    contextWindow: 262144,
+    pricing: cny(1.3, 6.5, 27),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  // The Pro/ and Qwen/ entries below were unpriced until 2026-08-03 (SiliconFlow's price
+  // list sits behind an authenticated console); prices below are its official CNY list
+  // prices.
+  {
+    modelId: "Pro/moonshotai/Kimi-K2.6",
+    displayName: "Kimi K2.6",
+    provider: "siliconflow",
+    contextWindow: 262144,
+    pricing: cny(1.1, 6.5, 27),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    // No cache-hit price on the list, so cache_read carries the input price.
+    modelId: "Qwen/Qwen3.6-35B-A3B",
+    displayName: "Qwen 3.6 35B A3B",
+    provider: "siliconflow",
+    contextWindow: 262144,
+    pricing: cny(1.8, 1.8, 10.8),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    // Added 2026-09-16, both ids confirmed in SiliconFlow's /v1/models listing. Hy4 preview is
+    // text only at CNY 6 input / 18 output / 0.3 cache hit, the list price as quoted again on
+    // 2026-10-10; siliconflow.cn/pricing shows the same input and output figures and no cache
+    // price. Its window is the 1,048,576 tokens that page now publishes for it (the row had
+    // borrowed the TokenDance row's 1,024,000). GLM-5.3 below is text only at 8 / 28 / 2, with
+    // the same 1M window as zai-org/GLM-5.2.
+    modelId: "tencent/Hy4-preview",
+    displayName: "Hy4 preview",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(0.3, 6, 18),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    modelId: "zai-org/GLM-5.3",
+    displayName: "GLM-5.3",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(2, 8, 28),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    // Added 2026-10-10 at the list price quoted that day: CNY 0.23 cache hit / 0.8 input / 2.8
+    // output, which is Z.AI's own CNY list for the model and what the TokenDance and Qwen
+    // Pay-As-You-Go rows store. siliconflow.cn/pricing currently shows it at CNY 0, beside the
+    // platform's free models; a free period is a rate no promotion field can hold, so the row
+    // keeps the list price. The window is the 1,048,576 tokens that page publishes. Natively
+    // multimodal (`vlm` there), and this group's openai-chat client forwards image_url parts.
+    modelId: "zai-org/GLM-5.3-Flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(0.23, 0.8, 2.8),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  {
+    modelId: "zai-org/GLM-5.2",
+    displayName: "GLM-5.2",
+    provider: "siliconflow",
+    contextWindow: 1048576,
+    pricing: cny(2, 8, 28),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: SILICONFLOW_BASE_URL
+  },
+  // -- Atlas Cloud (gateway: OpenAI-compatible protocol, preset base URL). Read 2026-10-11 from
+  // the public catalog API (GET https://api.atlascloud.ai/v1/models, no credential required):
+  // the ids exactly as listed, context windows, output limits, input modalities, quantization
+  // and prices. The API's `pricing` figures are USD per token; the pricing page
+  // (https://www.atlascloud.ai/pricing/models) states the same rates in USD per million
+  // tokens. The presets are DeepSeek and Qwen models the other gateway groups preset. For
+  // DeepSeek they are the releases TokenDance presets instead of the bare V4 ids (the user's
+  // call for TokenDance on 2026-10-10, made for this group on 2026-10-11), taken only where
+  // Atlas Cloud serves them at fp8: V4.1 Flash and V4 Pro 0813. Its V4 Flash 0731 is served at
+  // fp4, like the bare deepseek-v4-flash and deepseek-v4-pro, so none of the three is a preset.
+  // Qwen 3.6 35B A3B is the one the SiliconFlow and OpenRouter groups preset. Rows run newest
+  // release first within a series (V4.1 before V4). Neither source publishes a cache-write
+  // price for these rows, so cache_write carries the input price, as with the TokenDance and
+  // Qwen gateways below. Atlas Cloud bills both DeepSeek rows at one fixed rate
+  // (`pricing_mode` fixed, no time-based overrides), so neither declares DeepSeek's off-peak
+  // schedule. The DeepSeek rows have a 393,216-token output limit; Qwen 3.6 35B A3B, which
+  // takes image and video input, has 65,536.
+  {
+    // Image input, as the listing says (`text`, `image`).
+    modelId: "deepseek-ai/deepseek-v4.1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "atlascloud",
+    contextWindow: 1048576,
+    pricing: usd(0.03, 0.3, 1.2),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: ATLASCLOUD_BASE_URL
+  },
+  {
+    modelId: "deepseek-ai/deepseek-v4-pro-0813",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: "atlascloud",
+    contextWindow: 1048576,
+    pricing: usd(0.132, 1.32, 3.96),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: ATLASCLOUD_BASE_URL
+  },
+  {
+    // The API answers the rate billed today, USD 0.186 input / 1.11375 output, which the
+    // pricing page shows as a 25% model discount on a 0.248 / 1.485 list price. The row stores
+    // that list price and declares the promotion in `discount`, as every promoted row does.
+    // No cache-hit price is published (the API repeats the input price), so cache_read carries
+    // the input price too.
+    modelId: "qwen/qwen3.6-35b-a3b",
+    displayName: "Qwen 3.6 35B A3B",
+    provider: "atlascloud",
+    contextWindow: 262144,
+    pricing: usd(0.248, 0.248, 1.485),
+    discount: 0.25,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: ATLASCLOUD_BASE_URL
+  },
+  // -- TokenDance (gateway: OpenAI-compatible protocol, preset base URL). Re-read in full on
+  // 2026-10-10 from two public APIs that need no credential: the gateway's catalog (GET
+  // https://tokendance.space/gateway/v1/models — ids, context windows and supported protocols)
+  // and the portal's model list (GET https://tokendance.space/portal/api/models/all — names,
+  // input modalities and prices). Prices are the gateway's own CNY rates. The portal publishes
+  // what a model bills as `pricing.items` and, while a promotion runs, its list price beside
+  // that as `pricing.reference`. TokenDance charges no separate cache-write fee, so cache_write
+  // carries the input price. The per-model pages (tokendance.space/models/<id>) render in the
+  // browser from the same data.
+  //
+  // Discounts: every row here stores the official LIST price — the portal's `reference`, or
+  // its `items` where no promotion runs — the convention of the two Qwen groups below, and a
+  // promoted row declares its rate in `discount` rather than having its price rewritten: a
+  // promotion that lapses is then one field to delete, with the rate to return to still on the
+  // row. effectivePricing() applies it; a new Project stores the list price and the fraction as
+  // a promotion beside it, which the cost center applies, so it charges what the gateway
+  // charges. Eight rows carry a flat promotion as of 2026-10-10: ling-3.0-flash at 65% off, the
+  // three Doubao Seed rows (seed-2.1-pro, seed-2.1-turbo, seed-evolving) at 50%, glm-5.2 at 20%,
+  // and glm-5.3, glm-5.3-flash and qwen3.8-max at 10%; deepseek-v4.1-flash carries one on top
+  // of its schedule (below). The gateway also sells the DeepSeek V4 preview releases under the
+  // bare ids deepseek-v4-flash and deepseek-v4-pro ("… Preview"); they are deliberately not
+  // presets (the user's call, 2026-10-10): the dated rows below are the releases that
+  // superseded them.
+  //
+  // Three DeepSeek rows — deepseek-v4.1-flash, deepseek-v4-flash-0731 and deepseek-v4-pro-0813 —
+  // are priced by time instead (`pricing.time_pricing`): a peak price on weekdays 09:00-12:00
+  // and 14:00-18:00 Beijing time, DeepSeek's own windows, and half the peak list price at every
+  // other hour. They store the peak list price and declare DEEPSEEK_OFF_PEAK, as the direct
+  // DeepSeek rows do; so does the retired deepseek-v4-flash-vision-exp row, which is no longer a
+  // preset. On top of that schedule the gateway takes 20% off. On deepseek-v4.1-flash it does
+  // so in both tiers, and the row declares it in `discount` beside its schedule (re-read
+  // 2026-10-13). On the 0731 and 0813 rows it does so in the peak tier alone and bills
+  // off-peak at exactly half the list; a declared discount would come off the off-peak tier
+  // too, so theirs is not recorded, and the cost center prices their peak hours 25% above what
+  // the gateway bills. The rest of the group carries neither, so list price and billed rate
+  // coincide.
+  //
+  // A running promotion usually also shows in a model's `description`, which opens with a
+  // bracketed 限时 ("limited-time") tag. The tag is neither exhaustive nor authoritative on the
+  // rate: kimi-k3's still announces an offer that ended 2026-09-30, and the two Seed 2.1 rows'
+  // say 20% while their prices are at 50% off. The rows follow the prices.
+  //
+  // Display names follow the catalog's spelling of a model family where it has one (GLM-5.3,
+  // Qwen 3.8 Max, MiMo-V2.5), and otherwise the portal's name without its "Vendor: " prefix
+  // (Seed-2.1-Pro, Ling-3.1-flash, Step 5 Preview). Rows are in plain dictionary order by id. --
+  {
+    // Text only, 1,048,576 window. Priced by time (see the block comment): a CNY 3 input / 9
+    // output / 0.1 cache hit list, billed at 2.4 / 7.2 / 0.08 in the peak windows and at half
+    // the list, 1.5 / 4.5 / 0.05, at every other hour. It had been quoted on 2026-09-16 as a
+    // flat CNY 1.5 / 4.5 / 0.15 list at 10% off.
+    modelId: "deepseek-v4-flash-0731",
+    displayName: "DeepSeek V4 Flash 0731",
+    provider: "tokendance",
+    contextWindow: 1048576,
+    pricing: cny(0.1, 3, 9),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Retired from TokenDance's line-up 2026-09-16 (see ModelCatalogEntry.retired). It was sold
+    // on DeepSeek's own peak/off-peak schedule at the peak tier CNY 0.04 / 2 / 8, the price a
+    // sync keeps on the Projects still carrying it.
+    modelId: "deepseek-v4-flash-vision-exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL,
+    retired: true
+  },
+  {
+    // Text only, 1,000,000 window. Priced by time like the 0731 row: a CNY 9 input / 27 output
+    // / 0.3 cache hit list, DeepSeek's own peak price for V4 Pro 0813, billed at 7.2 / 21.6 /
+    // 0.24 in the peak windows and at half the list, 4.5 / 13.5 / 0.15, at every other hour. It
+    // had been quoted on 2026-09-16 as a flat CNY 4.5 / 13.5 / 0.45 list at 10% off.
+    modelId: "deepseek-v4-pro-0813",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.3, 9, 27),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Listed 2026-09-10; the dotted id is the seller's spelling of the model DeepSeek serves
+    // directly as `deepseek-flash`. 1M context and native multimodality. Its list follows the
+    // vendor's own schedule: the peak tier CNY 0.04 / 2 / 8, the direct row's figures, and half
+    // that outside Beijing weekday 09:00-12:00 and 14:00-18:00. The row stores the peak tier
+    // and declares DEEPSEEK_OFF_PEAK. The gateway also takes 20% off both tiers, declared in
+    // `discount`: it bills 0.032 / 1.6 / 6.4 at peak and 0.016 / 0.8 / 3.2 off-peak (see the
+    // block comment). Unlike the direct row this one needs no DeepSeek-specific pin — it is
+    // reached through this group's generic openai-chat client, which forwards image_url parts.
+    modelId: "deepseek-v4.1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    discount: 0.2,
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // The 1017 snapshot, priced exactly as the row above in the public catalog (re-read
+    // 2026-10-23): the same peak tier, schedule and 20% promotion.
+    modelId: "deepseek-v4.1-flash-1017",
+    displayName: "DeepSeek V4.1 Flash 1017",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    discount: 0.2,
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Dots Studio's (rednote / Xiaohongshu) Dots3-Note Preview: an open-weight MoE, 280B total
+    // and 16B active, the lightest of the Dots 3 family. TokenDance lists it as "Dots3-Note
+    // Preview（Free）": the words and the tag are kept as the seller spells them, as the
+    // OpenRouter `(free)` rows keep theirs, but the full-width parentheses are the typesetting
+    // of the seller's Chinese listing rather than part of the name, and a catalog name shows
+    // untranslated in every UI language — so it takes the ASCII " (" of every other
+    // parenthesised row, which reads right in both. The price is a
+    // genuine CNY 0 on every bucket, the same treatment as the OpenRouter `:free` rows, so
+    // costs compute to 0 and the free badge shows — over a 512,000-token context window, with
+    // openai:chat-completions and anthropic:messages as its supported_protocols (this group's
+    // openai-chat pin is a convention here, not the only shape the id serves). The gateway's
+    // own listing describes it as covering multimodal understanding, which is what the
+    // vision flag records; no image request was sent to it. Read 2026-09-15 from the
+    // gateway's /models listing and tokendance.space/models/dots-3-note-preview, and unchanged
+    // on 2026-10-10.
+    modelId: "dots-3-note-preview",
+    displayName: "Dots3-Note Preview (Free)",
+    provider: "tokendance",
+    contextWindow: 512e3,
+    pricing: cny(0, 0, 0),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Text only, 1,000,000 window. 20% off a CNY 8 input / 28 output / 2 cache hit list, the
+    // same list as glm-5.3, so the gateway bills 6.4 / 22.4 / 1.6.
+    modelId: "glm-5.2",
+    displayName: "GLM-5.2",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(2, 8, 28),
+    discount: 0.2,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    modelId: "glm-5.3",
+    displayName: "GLM-5.3",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(2, 8, 28),
+    discount: 0.1,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Natively multimodal per TokenDance's catalog entry, and this group's generic
+    // openai-chat client forwards image_url parts, so image input works on this path. Its
+    // supported_protocols are openai:chat-completions and anthropic:messages.
+    //
+    // The 50% promotion this row used to carry ran through 2026-09-09 24:00 (the catalog
+    // API's own description said so) and has been replaced by 10% off; the list price is
+    // unchanged, only the `discount` fraction moved.
+    modelId: "glm-5.3-flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.23, 0.8, 2.8),
+    discount: 0.1,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // A faster GLM-5.3 Flash (up to 200 tokens/s, per its listing), natively multimodal like
+    // it: text, image and video input, which this group's openai-chat client forwards.
+    // 1,000,000 window; CNY 2 input / 7 output / 0.57 cache hit, undiscounted.
+    modelId: "glm-5.3-flashx",
+    displayName: "GLM-5.3 FlashX",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.57, 2, 7),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // The same upstream model as the OpenRouter `tencent/hy4-preview` row above, reached
+    // through a second seller: each row records what its own seller charges, so the two must
+    // not be made to agree (the qwen3.8-flash pair below states the same rule). TokenDance
+    // sells it at CNY 6 input / 18 output / 0.3 cache hit, undiscounted, over a 1,024,000
+    // context window — both figures its own, neither copied from the OpenRouter listing.
+    //
+    // Text-only: the catalog entry advertises no image modality, matching the OpenRouter
+    // listing's text-in/text-out. Its supported_protocols are openai:chat-completions and
+    // openai:responses, so the openai-chat pin is this group's convention rather than the
+    // only shape the id serves.
+    modelId: "hy4-preview",
+    displayName: "Hy4 preview",
+    provider: "tokendance",
+    contextWindow: 1024e3,
+    pricing: cny(0.3, 6, 18),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // CNY 20 input / 100 output / 2 cache hit, Moonshot's own list, undiscounted as of
+    // 2026-10-10. The 40% promotion this row carried (seller's quote 2026-09-16, when the cache
+    // hit listed at 1.6) has ended; the description still opens with the tag of an offer that
+    // ran to 2026-09-30. Its supported_protocols is openai:chat-completions alone, so the
+    // openai-chat pin is the only shape this id serves.
+    modelId: "kimi-k3",
+    displayName: "Kimi K3",
+    provider: "tokendance",
+    contextWindow: 1048576,
+    pricing: cny(2, 20, 100),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Moonshot's Kimi K3.1 Pro (public catalog, read 2026-10-24): text, image and video input
+    // over a 1,000,000-token window, at CNY 13 input / 68 output / 1.3 cache hit, undiscounted
+    // and the same on both protocols the portal prices it on. Its supported_protocols are
+    // openai:chat-completions and anthropic:messages, so the openai-chat pin is the group's
+    // convention rather than the only shape the id serves; image parts go through this group's
+    // openai-chat client.
+    modelId: "kimi-k3.1-pro",
+    displayName: "Kimi K3.1 Pro",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(1.3, 13, 68),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // inclusionAI's Ling-3.0-flash: text only, with the 256,000-token window the listing gives
+    // (the model's native length; its description says it extends to 1M). 65% off a CNY 0.4
+    // input / 1.2 output / 0.08 cache hit list, so the gateway bills 0.14 / 0.42 / 0.028.
+    modelId: "ling-3.0-flash",
+    displayName: "Ling-3.0-flash",
+    provider: "tokendance",
+    contextWindow: 256e3,
+    pricing: cny(0.08, 0.4, 1.2),
+    discount: 0.65,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Text only, 1,000,000 window, at a CNY 0.4 input / 1.2 output / 0.08 cache hit list (the
+    // portal's `reference`, re-read 2026-10-14). Listed 2026-09-30 with a free trial served at
+    // 256K, which the portal bills at 0 until 2026-10-14 10:00. A free period is no `discount`,
+    // which is a fraction below 1. The row records the list price and the full window the
+    // model is sold at after the trial.
+    modelId: "ling-3.1-flash",
+    displayName: "Ling-3.1-flash",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.08, 0.4, 1.2),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Xiaomi's MiMo-V2.6 line, the three rows alike: text, image, audio and video input (image
+    // parts go through this group's openai-chat client), a 1,000,000 window, no promotion, and
+    // openai:chat-completions, openai:responses and anthropic:messages as supported_protocols.
+    // Flash at CNY 1 input / 2 output / 0.02 cache hit.
+    modelId: "mimo-v2.6-flash",
+    displayName: "MiMo-V2.6-Flash",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.02, 1, 2),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // Pro at CNY 3 input / 6 output / 0.025 cache hit.
+    modelId: "mimo-v2.6-pro",
+    displayName: "MiMo-V2.6-Pro",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.025, 3, 6),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // The Pro model served up to 20x faster, at ten times its price on every bucket: CNY 30
+    // input / 60 output / 0.25 cache hit (OpenRouter lists the same rates in USD for
+    // xiaomi/mimo-v2.6-pro-ultraspeed). The seller's id drops "pro"; its name keeps it.
+    modelId: "mimo-v2.6-ultraspeed",
+    displayName: "MiMo-V2.6-Pro-UltraSpeed",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.25, 30, 60),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // CNY 1 input / 4 output / 0.1 cache hit over a 1,000,000-token window, undiscounted (public
+    // catalog, re-read 2026-10-23). Pinned to Anthropic Messages: the gateway's Chat Completions
+    // stream for this model ends in a chunk without `delta`, which openai-chat cannot read
+    // (MMSP 0.5.5's registry pins the same).
+    modelId: "minimax-m3.1-flash",
+    displayName: "MiniMax M3.1 Flash",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.1, 1, 4),
+    supportsVision: true,
+    clientType: "ant-messages",
+    baseUrl: TOKENDANCE_MESSAGES_BASE_URL
+  },
+  {
+    // One model reached two ways, priced by whoever is selling it: this row carries the
+    // gateway's own CNY 0.8 input / 2.7 output / 0.1 cache hit, and the qwen-pay-as-you-go row
+    // carries Qwen's direct list price for the same id. The two figures agree since Qwen's
+    // 2026-09-16 re-pricing (its page had read 1 / 3 / 0.1), but only because the two sellers
+    // do: each row still follows its own seller.
+    //
+    // Its supported_protocols are openai:chat-completions, openai:responses and
+    // anthropic:messages, so the openai-chat pin here is the group's convention rather than the
+    // only shape the id serves. Natively multimodal per the catalog entry, and this group's
+    // openai-chat client forwards image_url parts, so image input works on this path.
+    //
+    // Undiscounted, so its list price and its billed rate coincide — unlike the
+    // qwen3.8-max row below, which is on 10% off.
+    modelId: "qwen3.8-flash",
+    displayName: "Qwen 3.8 Flash",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.1, 0.8, 2.7),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    modelId: "qwen3.8-max",
+    displayName: "Qwen 3.8 Max",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(1.5, 12, 36),
+    discount: 0.1,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // The three Doubao Seed rows share this note. Priced 2026-09-02 from the seller's quoted
+    // 50%-off rates, doubled back to the list price a row stores, and re-read 2026-10-10 from
+    // the portal, whose prices still put all three at 50% off a list of CNY 6 / 30 / 1.2
+    // (Pro, Evolving) and 3 / 15 / 0.6 (Turbo), input / output / cache hit — although the two
+    // 2.1 rows' descriptions announce 20%. Context windows and protocols from the catalog API,
+    // whose descriptions call the 2.1 models multimodal Coding/Agent models (this group's
+    // openai-chat client forwards image_url parts, so image input works on this path).
+    // seed-2.1-pro and seed-2.1-turbo also list openai:responses, so their openai-chat pin is
+    // the group's convention rather than the only shape they serve; seed-evolving lists
+    // chat-completions alone.
+    //
+    // Display names are the seller's own, as the catalog API spells them (re-read
+    // 2026-10-10): Seed-2.1-Pro, Seed-2.1-Turbo, Seed-Evolving.
+    modelId: "seed-2.1-pro",
+    displayName: "Seed-2.1-Pro",
+    provider: "tokendance",
+    contextWindow: 256e3,
+    pricing: cny(1.2, 6, 30),
+    discount: 0.5,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    modelId: "seed-2.1-turbo",
+    displayName: "Seed-2.1-Turbo",
+    provider: "tokendance",
+    contextWindow: 256e3,
+    pricing: cny(0.6, 3, 15),
+    discount: 0.5,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // A rolling id: the catalog API describes it as the newest Seed Coding/Agent model
+    // under one stable id — the same model as seed-2.1-pro at the time of reading — and
+    // the seller prices it the same.
+    modelId: "seed-evolving",
+    displayName: "Seed-Evolving",
+    provider: "tokendance",
+    contextWindow: 256e3,
+    pricing: cny(1.2, 6, 30),
+    discount: 0.5,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  {
+    // StepFun's Step 5 Preview: text, image and video input, a 1,000,000 window, and CNY 7
+    // input / 20 output / 0.35 cache hit, undiscounted.
+    modelId: "step-5-preview",
+    displayName: "Step 5 Preview",
+    provider: "tokendance",
+    contextWindow: 1e6,
+    pricing: cny(0.35, 7, 20),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: TOKENDANCE_BASE_URL
+  },
+  // -- Penguin Go. All models use MMSP HTTP; the SDK appends /stream to /api/v1.
+  // List prices are preset; authorization and sync supply the platform's promotions.
+  // The DeepSeek rows follow
+  // DeepSeek's own lineup (re-read 2026-09-16): V4.1 Flash as `deepseek-flash` and V4 Pro 0813
+  // as `deepseek-v4-pro`; the two retired V4 Flash ids, served from V4.1 Flash, are not resold.
+  // Gemini 3.7 Flash is no longer preset (2026-10-10): Google retires the id and reroutes it to
+  // 3.8 Flash.
+  {
+    // Owner-confirmed relay list prices (2026-10-14), matching Google's introductory
+    // rates. Any further Penguin Go promotion comes from authorization or model sync.
+    modelId: "gemini-4-argon",
+    displayName: "Gemini 4 Argon",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.1, 2, 10),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-3.8-flash",
+    displayName: "Gemini 3.8 Flash",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.075, 0.75, 3.75),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-3.6-flash",
+    displayName: "Gemini 3.6 Flash",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.075, 0.75, 3.75),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-3.5-flash",
+    displayName: "Gemini 3.5 Flash",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 9),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-3.5-flash-lite",
+    displayName: "Gemini 3.5 Flash-Lite",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.03, 0.3, 2.5),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-3.1-flash-lite",
+    displayName: "Gemini 3.1 Flash-Lite",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.025, 0.25, 1.5),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-3.1-pro-preview",
+    displayName: "Gemini 3.1 Pro (Preview)",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1048576,
+    pricing: usd(0.2, 2, 12),
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "deepseek-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1e6,
+    pricing: cny(0.04, 2, 8),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  {
+    modelId: "deepseek-v4-pro",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: PENGUIN_GO_PROVIDER_ID,
+    contextWindow: 1e6,
+    pricing: cny(0.3, 9, 27),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    clientType: "mmsp",
+    baseUrl: PENGUIN_GO_BASE_URL
+  },
+  // -- OpenCode Go. Latest generation of each model tier from opencode.ai/v2/docs/console/go/
+  // (2026-10-14); context and vision from OpenCode's models.dev registry. The Code, Pro,
+  // Flash, Max and Plus tiers are independent: a newer Flash does not replace a Pro.
+  // Generic clients pin the documented wire protocol. The /zen/go aliases remain live;
+  // keeping them avoids changing existing group endpoints. Requests name their conversation
+  // through the x-opencode-session header core's GenerativeModel attaches.
+  // Prices measure subscription allowance consumed, not the monthly fee. Cache writes use
+  // the input price when no write premium is published. GPT Luna (>272K), Haiku (>100K),
+  // Grok (>200K) and Qwen Plus (>256K) store the base tier. DeepSeek stores its peak rates
+  // and DEEPSEEK_OFF_PEAK schedule. Allowance multipliers are not price discounts.
+  // Superseded rows remain retired for saved Projects and historical off-peak pricing;
+  // removal follows ModelCatalogEntry.retired and the backward-compatibility changelog.
+  {
+    modelId: "cat-jam-free",
+    displayName: "Cat Jam Free",
+    provider: "opencode-go",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "claude-haiku-5-5",
+    displayName: "Claude Haiku 5.5",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.01, 0.125, 0.5),
+    supportsVision: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "deepseek-v4.1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(6e-3, 0.3, 1.2),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "deepseek-v4-flash",
+    displayName: "DeepSeek V4 Flash",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(6e-3, 0.3, 1.2),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "deepseek-v4-flash-vision-exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(6e-3, 0.3, 1.2),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: true,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "deepseek-v4-pro",
+    displayName: "DeepSeek V4 Pro",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.044, 1.32, 3.96),
+    offPeakDiscount: DEEPSEEK_OFF_PEAK,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "gemini-4-argon-contributor",
+    displayName: "Gemini 4 Argon Contributor",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(0.2, 4, 20),
+    supportsVision: true,
+    clientType: "google-genai",
+    baseUrl: OPENCODE_GO_GOOGLE_BASE_URL
+  },
+  {
+    modelId: "glm-5.3",
+    displayName: "GLM-5.3",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "glm-5.3-flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.03, 0.15, 0.5),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "glm-5.2",
+    displayName: "GLM-5.2",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "glm-5.1",
+    displayName: "GLM-5.1",
+    provider: "opencode-go",
+    contextWindow: 202752,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "gpt-6-luna",
+    displayName: "GPT-6 Luna",
+    provider: "opencode-go",
+    contextWindow: 105e4,
+    pricing: usd(0.01, 0.125, 0.5),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "gpt-5.6-luna",
+    displayName: "GPT-5.6 Luna",
+    provider: "opencode-go",
+    contextWindow: 105e4,
+    pricing: usd(0.02, 0.25, 1.2),
+    supportsVision: true,
+    retired: true,
+    clientType: "openai-responses",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "grok-4.7",
+    displayName: "Grok 4.7",
+    provider: "opencode-go",
+    contextWindow: 5e5,
+    pricing: usd(0.5, 2, 6),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "grok-4.6",
+    displayName: "Grok 4.6",
+    provider: "opencode-go",
+    contextWindow: 5e5,
+    pricing: usd(0.5, 2, 6),
+    supportsVision: true,
+    retired: true,
+    clientType: "openai-responses",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "hy4-preview",
+    displayName: "Hy4 preview",
+    provider: "opencode-go",
+    contextWindow: 1024e3,
+    pricing: usd(0.042, 0.834, 2.501),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    // models.dev lists a 256,000-token context for Hy3 but caps its input at 192,000. Compaction
+    // derives from this field, so it records the input cap: at 256,000 a Session would keep
+    // sending prompts the upstream refuses until it compacted at ~254K.
+    modelId: "hy3",
+    displayName: "Hy3",
+    provider: "opencode-go",
+    contextWindow: 192e3,
+    pricing: usd(0.035, 0.14, 0.58),
+    supportsVision: false,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "kimi-k3",
+    displayName: "Kimi K3",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(0.3, 3, 15),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "kimi-k2.7-code",
+    displayName: "Kimi K2.7 Code",
+    provider: "opencode-go",
+    contextWindow: 262144,
+    pricing: usd(0.19, 0.95, 4),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "kimi-k2.6",
+    displayName: "Kimi K2.6",
+    provider: "opencode-go",
+    contextWindow: 262144,
+    pricing: usd(0.16, 0.95, 4),
+    supportsVision: true,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "longcat-2.5-preview-free",
+    displayName: "LongCat 2.5 Preview Free",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "longcat-2.0",
+    displayName: "LongCat 2.0",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(6e-3, 0.3, 1.2),
+    supportsVision: false,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "mimo-v2.6-flash",
+    displayName: "MiMo-V2.6-Flash",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(28e-4, 0.14, 0.28),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "mimo-v2.6-pro",
+    displayName: "MiMo-V2.6-Pro",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(3625e-6, 0.435, 0.87),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "mimo-v2.5",
+    displayName: "MiMo-V2.5",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(28e-4, 0.14, 0.28),
+    supportsVision: true,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "mimo-v2.5-pro",
+    displayName: "MiMo-V2.5-Pro",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(3625e-6, 0.435, 0.87),
+    supportsVision: false,
+    retired: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "minimax-m3",
+    displayName: "MiniMax M3",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.06, 0.3, 1.2),
+    supportsVision: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "minimax-m2.7",
+    displayName: "MiniMax M2.7",
+    provider: "opencode-go",
+    contextWindow: 204800,
+    pricing: usd(0.06, 0.375, 1.2),
+    supportsVision: false,
+    retired: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "muse-spark-1.3-contributor",
+    displayName: "Muse Spark 1.3 Contributor",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(2e-3, 0.1, 0.2),
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "muse-spark-1.2-contributor",
+    displayName: "Muse Spark 1.2 Contributor",
+    provider: "opencode-go",
+    contextWindow: 1048576,
+    pricing: usd(2e-3, 0.1, 0.2),
+    supportsVision: true,
+    retired: true,
+    clientType: "openai-responses",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "qwen3.8-flash",
+    displayName: "Qwen 3.8 Flash",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.016, 0.2, 0.47),
+    supportsVision: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "qwen3.8-max",
+    displayName: "Qwen 3.8 Max",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.25, 2.5, 6),
+    supportsVision: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "qwen3.7-max",
+    displayName: "Qwen 3.7 Max",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 3.125, 7.5),
+    supportsVision: false,
+    retired: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "qwen3.7-plus",
+    displayName: "Qwen 3.7 Plus",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.04, 0.5, 1.6),
+    supportsVision: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    modelId: "qwen3.6-plus",
+    displayName: "Qwen 3.6 Plus",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0.05, 0.625, 3),
+    supportsVision: true,
+    retired: true,
+    clientType: "ant-messages",
+    baseUrl: OPENCODE_GO_MESSAGES_BASE_URL
+  },
+  {
+    // The context spans 1M tokens, but models.dev caps input at 524,288; compaction uses input.
+    modelId: "space-bunny",
+    displayName: "Space Bunny",
+    provider: "opencode-go",
+    contextWindow: 524288,
+    pricing: usd(0.03, 0.15, 0.6),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  {
+    modelId: "step-5-preview-free",
+    displayName: "Step 5 Preview Free",
+    provider: "opencode-go",
+    contextWindow: 1e6,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENCODE_GO_BASE_URL
+  },
+  // -- Qwen Token Plan (subscription gateway; vision flags per the plan's supported-model
+  // table, and for the rows added 2026-09-16 per their model pages' input modalities).
+  // Pricing and context windows from each model's page at www.qianwenai.com/models/<id>
+  // (official CNY list prices; limited-time promotions such as the 20%/50% off discounts are
+  // not stored). Lineup updated 2026-09-16: deepseek-v4.1-flash, deepseek-v4-pro-0813,
+  // glm-5.3 and qwen3.8-flash join; deepseek-v4-flash-0731, deepseek-v4-pro and glm-5.2
+  // leave the plan.
+  //
+  // The two DeepSeek rows are priced by time of day: their pages list a peak and an off-peak
+  // price, the off-peak one exactly half on every bucket and in force from 22:00 to 08:00
+  // Beijing time every day. They store the PEAK price and declare QWEN_OFF_PEAK, the same
+  // shape the DeepSeek rows take with DeepSeek's own schedule. --
+  {
+    // Peak CNY 2 input / 8 output / 0.2 cache hit; off-peak 1 / 4 / 0.1. Text and image input.
+    modelId: "deepseek-v4.1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "qwen-token-plan",
+    contextWindow: 1e6,
+    pricing: cny(0.2, 2, 8),
+    offPeakDiscount: QWEN_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  {
+    // Peak CNY 9 input / 27 output / 0.9 cache hit; off-peak 4.5 / 13.5 / 0.45. Text only.
+    modelId: "deepseek-v4-pro-0813",
+    displayName: "DeepSeek V4 Pro 0813",
+    provider: "qwen-token-plan",
+    contextWindow: 1e6,
+    pricing: cny(0.9, 9, 27),
+    offPeakDiscount: QWEN_OFF_PEAK,
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  {
+    modelId: "glm-5.3",
+    displayName: "GLM-5.3",
+    provider: "qwen-token-plan",
+    contextWindow: 1048576,
+    pricing: cny(2, 8, 28),
+    supportsVision: false,
+    clientType: "openai-chat",
+    baseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  {
+    modelId: "qwen3.8-flash",
+    displayName: "Qwen 3.8 Flash",
+    provider: "qwen-token-plan",
+    contextWindow: 1e6,
+    pricing: cny(0.1, 0.8, 2.7),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  {
+    modelId: "qwen3.8-max",
+    displayName: "Qwen 3.8 Max",
+    provider: "qwen-token-plan",
+    contextWindow: 1e6,
+    pricing: cny(1.5, 12, 36),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  {
+    modelId: "qwen3.7-plus",
+    displayName: "Qwen 3.7 Plus",
+    provider: "qwen-token-plan",
+    contextWindow: 1e6,
+    pricing: cny(0.4, 2, 8),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_TOKEN_PLAN_BASE_URL
+  },
+  // -- Qwen Pay-As-You-Go (DashScope's OpenAI-compatible pay-per-token marketplace; official
+  // CNY list prices and specs from each model's page at www.qianwenai.com/models/<id> —
+  // resold third-party models keep their upstream ids exactly as the page lists them: kimi/
+  // and ZHIPU/ carry vendor prefixes, DeepSeek is listed bare). Lineup updated 2026-09-16:
+  // deepseek-v4.1-flash, kimi/kimi-k2.8-preview and ZHIPU/GLM-5.3-Flash join;
+  // deepseek-v4-flash-0731 and ZHIPU/GLM-5.2 leave. The DeepSeek row is priced by time of day
+  // on QWEN_OFF_PEAK and stores its peak price, as in the Token Plan group above. --
+  {
+    // Peak CNY 2 input / 8 output / 0.2 cache hit; off-peak (22:00-08:00 Beijing) 1 / 4 / 0.1.
+    // Text and image input, which this group's openai-chat client forwards.
+    modelId: "deepseek-v4.1-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1e6,
+    pricing: cny(0.2, 2, 8),
+    offPeakDiscount: QWEN_OFF_PEAK,
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    modelId: "kimi/kimi-k3",
+    displayName: "Kimi K3",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1048576,
+    pricing: cny(2, 20, 100),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    // CNY 6.5 input / 27 output / 1.7 cache hit over a 1M window (1,048,576, as on the
+    // kimi/kimi-k3 row); text, image and video input.
+    modelId: "kimi/kimi-k2.8-preview",
+    displayName: "Kimi K2.8 Preview",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1048576,
+    pricing: cny(1.7, 6.5, 27),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    // Official CNY list price from www.qianwenai.com/models/qwen3.8-flash, re-read 2026-09-16:
+    // CNY 0.8 input / CNY 0.1 cache hit / CNY 2.7 output per MTok (it had been 1 / 0.1 / 3),
+    // over a 1M-token input window with a 131K output cap. Its input modalities include
+    // images and video, and this group's openai-chat client converts image parts.
+    modelId: "qwen3.8-flash",
+    displayName: "Qwen 3.8 Flash",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1e6,
+    pricing: cny(0.1, 0.8, 2.7),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    modelId: "qwen3.8-max",
+    displayName: "Qwen 3.8 Max",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1e6,
+    pricing: cny(1.5, 12, 36),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    modelId: "qwen3.7-plus",
+    displayName: "Qwen 3.7 Plus",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1e6,
+    pricing: cny(0.4, 2, 8),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  {
+    // CNY 0.8 input / 2.8 output / 0.23 cache hit over a 1M window with a 131K output cap;
+    // natively multimodal like every other GLM-5.3 Flash row, and this group's openai-chat
+    // client forwards image parts.
+    modelId: "ZHIPU/GLM-5.3-Flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "qwen-pay-as-you-go",
+    contextWindow: 1e6,
+    pricing: cny(0.23, 0.8, 2.8),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: QWEN_PAYG_BASE_URL
+  },
+  // -- ModelScope (api-inference aggregate, preset base URL). Preset ids keep ModelScope's
+  // own `org/Model-Name` spelling; DeepSeek was read from the endpoint's public /v1/models
+  // listing (2026-09-18), and the Qwen additions from their public model pages (2026-09-20).
+  //
+  // ModelScope serves all three presets through its OpenAI Responses endpoint, so every row pins
+  // MMSP's generic Responses client. Keeping the protocol explicit also prevents ids from
+  // auto-routing to vendor-specific clients with a different request shape.
+  //
+  // The window and vision flags are NOT read from ModelScope's own docs — its model pages are
+  // client-rendered and carry no price. They repeat what other rows of the same models record,
+  // because ModelScope serves the same upstream ids.
+  //
+  // `pricing` is deliberately absent. ModelScope bills for api-inference and does not publish a
+  // read-able rate, so an absent pricing — which the catalog reads as "nobody has looked this
+  // up", leaving the row unbadged and the group's usage uncosted — is the honest state. A
+  // three-zero pricing would be worse than absent here, not better: `isFreeModel` would badge a
+  // billed gateway as "Free". Whoever next reads ModelScope's pricing page should fill it in. --
+  {
+    modelId: "deepseek-ai/DeepSeek-V4.1-Flash",
+    displayName: "DeepSeek V4.1 Flash",
+    provider: MODELSCOPE_PROVIDER_ID,
+    contextWindow: 1e6,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: MODELSCOPE_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.8-27B",
+    displayName: "Qwen 3.8 27B",
+    provider: MODELSCOPE_PROVIDER_ID,
+    contextWindow: 262144,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: MODELSCOPE_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.8-Flash-Next",
+    displayName: "Qwen 3.8 Flash Next",
+    provider: MODELSCOPE_PROVIDER_ID,
+    contextWindow: 262144,
+    supportsVision: true,
+    clientType: "openai-responses",
+    baseUrl: MODELSCOPE_BASE_URL
+  },
+  // -- MiniMax (official USD pay-as-you-go list prices, standard tier at <=512K input — every
+  // rate doubles above 512K, and the priority tier is 1.5x). --
+  {
+    // Pay-as-you-go list price, one tier (MiniMax pricing guide, read 2026-10-19 by MMSP 0.5.5's
+    // registry): $0.15 input / $0.60 output / $0.015 cache hit over a 1,000,000-token window.
+    modelId: "MiniMax-M3.1-Flash",
+    displayName: "MiniMax M3.1 Flash",
+    provider: "minimax",
+    contextWindow: 1e6,
+    pricing: usd(0.015, 0.15, 0.6),
+    supportsVision: true
+  },
+  {
+    modelId: "MiniMax-M3",
+    displayName: "MiniMax M3",
+    provider: "minimax",
+    contextWindow: 1e6,
+    pricing: usd(0.06, 0.3, 1.2),
+    supportsVision: true
+  },
+  // -- Google Gemini (official USD pricing). Gemini 3.7 Flash left the group on 2026-10-10:
+  // Google retires the id and reroutes it to 3.8 Flash, at the same price. --
+  {
+    // ai.google.dev/gemini-api/docs/pricing#gemini-4-argon (2026-10-14): standard list
+    // prices with a 50% introductory discount through 2026-11-30. Only implicit caching
+    // is supported; cache_write is the ordinary input price.
+    modelId: "gemini-4-argon",
+    displayName: "Gemini 4 Argon",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.2, 4, 20),
+    discount: 0.5,
+    supportsVision: true
+  },
+  {
+    // Same list price and same launch discount as the gemini-3.6-flash row below: Google
+    // halves all three rates through 2026-12-31. Like that row this one declares the
+    // promotion in `discount`, so the list price stays on file — here and in a Project, which
+    // is seeded with the fraction beside it — while the cost center bills the 0.075/0.75/3.75
+    // Google actually charges today. One field to delete when the promotion lapses.
+    modelId: "gemini-3.8-flash",
+    displayName: "Gemini 3.8 Flash",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 7.5),
+    discount: 0.5,
+    supportsVision: true
+  },
+  {
+    // Same list price and same launch discount as the 3.8 row: Google halves all three rates
+    // through 2026-12-31 (Google's pricing page, read 2026-09-09).
+    modelId: "gemini-3.6-flash",
+    displayName: "Gemini 3.6 Flash",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 7.5),
+    discount: 0.5,
+    supportsVision: true
+  },
+  {
+    // Google's 3.5 tier list price ($1.50 / $9.00 / $0.15 cache hit), which carries no launch
+    // discount — re-read 2026-09-09. The $9 output is what the tier costs, not a promotion.
+    modelId: "gemini-3.5-flash",
+    displayName: "Gemini 3.5 Flash",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.15, 1.5, 9),
+    supportsVision: true
+  },
+  {
+    modelId: "gemini-3.5-flash-lite",
+    displayName: "Gemini 3.5 Flash-Lite",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.03, 0.3, 2.5),
+    supportsVision: true
+  },
+  {
+    modelId: "gemini-3.1-flash-lite",
+    displayName: "Gemini 3.1 Flash-Lite",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.025, 0.25, 1.5),
+    supportsVision: true
+  },
+  {
+    // ≤200K input tier; >200K has official surcharge pricing (see file header comment).
+    modelId: "gemini-3.1-pro-preview",
+    displayName: "Gemini 3.1 Pro (Preview)",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.2, 2, 12),
+    supportsVision: true
+  },
+  {
+    modelId: "gemini-3-flash-preview",
+    displayName: "Gemini 3 Flash (Preview)",
+    provider: "google",
+    contextWindow: 1048576,
+    pricing: usd(0.05, 0.5, 3),
+    supportsVision: true
+  },
+  // -- Anthropic (official USD pricing; cache write = 1.25 x input, the 5-minute write).
+  // Re-read 2026-08-20 from platform.claude.com/docs/en/about-claude/pricing; the 5.5 / 5.1
+  // generation added 2026-10-10 from MMSP 0.5.2's registry and the copy of that page it
+  // carries (llmsdk_docs/claude5_5/docs/pricing.md). Sonnet 5's $2 input / $10 output is its
+  // standard rate rather than an introductory one, so it prices below Sonnet 4.6 — that
+  // inversion is Anthropic's list, not a transcription slip. The cache-hit multiplier is no
+  // longer one figure: 0.025x input on Fable 5.1, 0.05x on Opus 5.5 and Sonnet 5.5, 0.1x on
+  // every other row. Anthropic bills the full 1M window at a single rate on every row but
+  // Haiku 5.5, which the file header's tier list names; and the fast-mode premium on Opus 5.5
+  // ($8 input / $40 output) and Opus 5 / Opus 4.8 ($10 / $50) is a separate tier these rows do
+  // not record. Every row takes image input over a 1,000,000-token window. --
+  {
+    // $10 input / $12.50 cache write / $0.25 cache hit / $50 output.
+    modelId: "claude-fable-5-1",
+    displayName: "Claude Fable 5.1",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.25, 12.5, 50),
+    supportsVision: true
+  },
+  {
+    modelId: "claude-fable-5",
+    displayName: "Claude Fable 5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(1, 12.5, 50),
+    supportsVision: true
+  },
+  {
+    // $0.10 input / $0.125 cache write / $0.01 cache hit / $0.50 output for a prompt of up to
+    // 100,000 tokens, counting cache reads and writes; a longer prompt pays 5x on every rate,
+    // for the whole request. The base tier is recorded. MMSP 0.5.2 is the first release whose
+    // official Anthropic client knows the model: it runs `none` thinking at low effort, refuses
+    // a forced tool choice, which Haiku 5.5 answers without thinking, and refuses fast mode,
+    // which Haiku 5.5 rejects (fastModeProtocol offers no toggle for it either).
+    modelId: "claude-haiku-5-5",
+    displayName: "Claude Haiku 5.5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.01, 0.125, 0.5),
+    supportsVision: true
+  },
+  {
+    // $4 input / $5 cache write / $0.20 cache hit / $20 output.
+    modelId: "claude-opus-5-5",
+    displayName: "Claude Opus 5.5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.2, 5, 20),
+    supportsVision: true
+  },
+  {
+    modelId: "claude-opus-5",
+    displayName: "Claude Opus 5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 6.25, 25),
+    supportsVision: true
+  },
+  {
+    modelId: "claude-opus-4-8",
+    displayName: "Claude Opus 4.8",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 6.25, 25),
+    supportsVision: true
+  },
+  {
+    modelId: "claude-opus-4-7",
+    displayName: "Claude Opus 4.7",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.5, 6.25, 25),
+    supportsVision: true
+  },
+  {
+    // $2 input / $2.50 cache write / $0.10 cache hit / $10 output. MMSP 0.5.0 listed the cache
+    // hit at $0.20, Sonnet 5's figure; 0.5.2 lowered it to the 0.05x of input that Anthropic's
+    // pricing page states for this model.
+    modelId: "claude-sonnet-5-5",
+    displayName: "Claude Sonnet 5.5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.1, 2.5, 10),
+    supportsVision: true
+  },
+  {
+    modelId: "claude-sonnet-5",
+    displayName: "Claude Sonnet 5",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.2, 2.5, 10),
+    supportsVision: true
+  },
+  {
+    modelId: "claude-sonnet-4-6",
+    displayName: "Claude Sonnet 4.6",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    pricing: usd(0.3, 3.75, 15),
+    supportsVision: true
+  },
+  // -- OpenAI (official USD pricing) --
+  {
+    // OpenAI's list price per MMSP 0.5.2's registry and the model page it carries
+    // (llmsdk_docs/gpt6/docs/gpt-6.1-sol.md, read 2026-10-10): $2 input / $0.10 cached input
+    // (5% of input) / $2.50 cache write / $10 output, over a 1,050,000-token window with text
+    // and image input. As on gpt-6-astra below, cache_write carries the published cache-write
+    // price; a prompt above 272K tokens pays 2x input and cache and 1.5x output for the whole
+    // request, a tier this row does not record.
+    modelId: "gpt-6.1-sol",
+    displayName: "GPT-6.1 Sol",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.1, 2.5, 10),
+    supportsVision: true
+  },
+  {
+    // OpenAI's list price (developers.openai.com/api/docs/models/gpt-6-astra, read
+    // 2026-09-09) is $10 input / $1 cached input / $12.5 cache write / $50 output. Per the
+    // bucket convention at the top of this file, cache_write carries the published
+    // cache-write price of 12.5; the $10 rate applies only to input that is not written to
+    // cache, a split the three buckets do not express. Every rate doubles above 272K input
+    // tokens (output 1.5x) — the base tier is what this row records, as the header says.
+    modelId: "gpt-6-astra",
+    displayName: "GPT-6 Astra",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(1, 12.5, 50),
+    supportsVision: true
+  },
+  {
+    // https://developers.openai.com/api/docs/models/gpt-6-luna (2026-10-14).
+    // Base tier: above 272K input, input/cache rates double and output is 1.5x.
+    modelId: "gpt-6-luna",
+    displayName: "GPT-6 Luna",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.01, 0.125, 0.5),
+    supportsVision: true
+  },
+  {
+    // The bare gpt-5.6 id routes to gpt-5.6-sol upstream and is priced as that tier, so the
+    // row names the Sol codename its siblings and the openai/gpt-5.6-sol row already show —
+    // the id stays bare, only the label says which variant this is. This row and the two
+    // gpt-5.6 rows below mirror the
+    // openai/gpt-5.6-* OpenRouter rows above, which carry the gateway's (currently
+    // discounted) rates instead of this list price.
+    modelId: "gpt-5.6",
+    displayName: "GPT-5.6 Sol",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.5, 5, 30),
+    supportsVision: true
+  },
+  {
+    modelId: "gpt-5.6-luna",
+    displayName: "GPT-5.6 Luna",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.02, 0.2, 1.2),
+    supportsVision: true
+  },
+  {
+    modelId: "gpt-5.6-terra",
+    displayName: "GPT-5.6 Terra",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.2, 2, 12),
+    supportsVision: true
+  },
+  {
+    modelId: "gpt-5.5",
+    displayName: "GPT-5.5",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.5, 5, 30),
+    supportsVision: true
+  },
+  {
+    // No official cache discount: cache_read uses the standard input price.
+    modelId: "gpt-5.5-pro",
+    displayName: "GPT-5.5 Pro",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(30, 30, 180),
+    supportsVision: true
+  },
+  {
+    modelId: "gpt-5.4",
+    displayName: "GPT-5.4",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(0.25, 2.5, 15),
+    supportsVision: true
+  },
+  {
+    modelId: "gpt-5.4-mini",
+    displayName: "GPT-5.4 mini",
+    provider: "openai",
+    contextWindow: 4e5,
+    pricing: usd(0.075, 0.75, 4.5),
+    supportsVision: true
+  },
+  {
+    modelId: "gpt-5.4-nano",
+    displayName: "GPT-5.4 nano",
+    provider: "openai",
+    contextWindow: 4e5,
+    pricing: usd(0.02, 0.2, 1.25),
+    supportsVision: true
+  },
+  {
+    // No official cache discount: cache_read uses the standard input price.
+    modelId: "gpt-5.4-pro",
+    displayName: "GPT-5.4 Pro",
+    provider: "openai",
+    contextWindow: 105e4,
+    pricing: usd(30, 30, 180),
+    supportsVision: true
+  },
+  // -- Z.AI (GLM) --
+  {
+    // Announced 2026-08-14. Z.AI's price
+    // list (docs.z.ai/guides/overview/pricing, read 2026-08-18) publishes the same USD rates
+    // as glm-5.2 / glm-5.1.
+    modelId: "glm-5.3",
+    displayName: "GLM-5.3",
+    provider: "zhipu",
+    contextWindow: 1e6,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false
+  },
+  {
+    // Z.AI's price list (docs.z.ai/guides/overview/pricing) publishes $0.15 input / $0.03
+    // cached input / $0.50 output; a 50% promotion halved all three through 24:00 on
+    // 2026-09-09 (UTC+8). Direct-vendor rows record the vendor's list price, so that is what
+    // is stored here — the OpenRouter z-ai/glm-5.3-flash row above carries the rate its
+    // default endpoint bills, which runs a promotion of its own.
+    //
+    // The model is natively multimodal (docs.z.ai/guides/vlm/glm-5.3-flash: images, video
+    // and files), and it is the one GLM id whose images MMSP's Z.AI client forwards — as
+    // image_url parts, in a prompt and in a tool result alike. Every other GLM id refuses
+    // one outright ("GLM <id> does not support image inputs."), which is why the rest of
+    // this group is vision-off.
+    modelId: "glm-5.3-flash",
+    displayName: "GLM-5.3 Flash",
+    provider: "zhipu",
+    contextWindow: 1e6,
+    pricing: usd(0.03, 0.15, 0.5),
+    supportsVision: true
+  },
+  {
+    modelId: "glm-5.2",
+    displayName: "GLM-5.2",
+    provider: "zhipu",
+    contextWindow: 1e6,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false
+  },
+  {
+    modelId: "glm-5.1",
+    displayName: "GLM-5.1",
+    provider: "zhipu",
+    contextWindow: 2e5,
+    pricing: usd(0.26, 1.4, 4.4),
+    supportsVision: false
+  },
+  {
+    modelId: "glm-5",
+    displayName: "GLM-5",
+    provider: "zhipu",
+    contextWindow: 2e5,
+    pricing: usd(0.2, 1, 3.2),
+    supportsVision: false
+  },
+  // -- Moonshot (Kimi) (official CNY pricing) --
+  {
+    // Moonshot's list (pricing page, read 2026-10-23 for MMSP 0.5.6's registry): CNY 13 input /
+    // 68 output / 1.3 cache hit over a 1,048,576-token window, image input. The default
+    // five-minute cache write bills at the input price, which is what cache_write stores; the
+    // one-hour write, which only MMSP's `prompt_caching` enhance asks for, bills twice that.
+    modelId: "kimi-k3.1-pro",
+    displayName: "Kimi K3.1 Pro",
+    provider: "moonshot",
+    contextWindow: 1048576,
+    pricing: cny(1.3, 13, 68),
+    supportsVision: true
+  },
+  {
+    // The same model at two to three times the output speed (Moonshot's model list), at twice
+    // the input, output and cache-write prices and the same cache hit: CNY 26 / 136 / 1.3.
+    // Moonshot's vision guide names only kimi-k3.1-pro, but its model list files both ids under
+    // multimodal models, and MMSP's live probe read a base64 image through this one.
+    modelId: "kimi-k3.1-pro-highspeed",
+    displayName: "Kimi K3.1 Pro Highspeed",
+    provider: "moonshot",
+    contextWindow: 1048576,
+    pricing: cny(1.3, 26, 136),
+    supportsVision: true
+  },
+  {
+    modelId: "kimi-k3",
+    displayName: "Kimi K3",
+    provider: "moonshot",
+    contextWindow: 1048576,
+    pricing: cny(2, 20, 100),
+    supportsVision: true
+  },
+  {
+    modelId: "kimi-k2.6",
+    displayName: "Kimi K2.6",
+    provider: "moonshot",
+    contextWindow: 262144,
+    pricing: cny(1.1, 6.5, 27),
+    supportsVision: true
+  },
+  {
+    modelId: "kimi-k2.5",
+    displayName: "Kimi K2.5",
+    provider: "moonshot",
+    contextWindow: 262144,
+    pricing: cny(0.7, 4, 21),
+    supportsVision: true
+  },
+  // -- vLLM (self-hosted: the models MMSP's openai-chat-vllm-adapter client carries a
+  // per-model thinking switch for, as published at recipes.vllm.ai — read 2026-09-03).
+  //
+  // Every row prices at zero, and omits two other things, all because the user runs the server:
+  // - **zero pricing**. There is no seller charging per token: what the deployment costs is
+  //   the operator's own hardware, which no catalog rate could express. Three zero buckets are
+  //   the same genuine $0 tier the `:free` gateway rows carry, so these models show the free
+  //   badge and contribute 0 to the cost center rather than the "unpriced" mark — which is the
+  //   truthful reading of a self-hosted endpoint that bills nobody.
+  // - **the local server's base URL** (VLLM_LOCAL_BASE_URL), where `vllm serve` answers when
+  //   it runs on the same machine with its default port. A new Project stores it once, on the
+  //   group's table; a deployment elsewhere changes the group's base URL.
+  // - **no auto-routing**. Each row pins openai-chat-vllm-adapter explicitly, and the pin is
+  //   load-bearing twice over: `Qwen/*` begins with no family AutoLLMClient knows and would
+  //   be rejected outright, while `deepseek-ai/DeepSeek-V4-*` begins with "deepseek-" and
+  //   would reach DeepSeek's official Responses client — pointed at a vLLM server.
+  //
+  // contextWindow is the recipe's NATIVE length, which is the most a deployment can serve
+  // without reconfiguration; an operator may serve less (`--max-model-len` below the native
+  // limit) or, for the Qwen rows, far more with YaRN rope scaling. The catalog cannot know
+  // which, and it derives the compaction thresholds from this number, so the honest default
+  // is the checkpoint's own figure — an entry without one would be assumed to be 128000.
+  // The DeepSeek rows additionally document `--max-model-len >= 393216` as the floor for
+  // their top reasoning levels, which is well inside the window recorded here.
+  {
+    modelId: "deepseek-ai/DeepSeek-V4-Flash",
+    displayName: "DeepSeek V4 Flash",
+    provider: "vllm",
+    contextWindow: 1e6,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    // The experimental vision revision: DeepSeek's first multimodal V4, a ViT tower on the
+    // same language backbone. Its recipe verifies a 32K deployment and notes the 1M the
+    // checkpoint advertises was not what was measured; the window below is the checkpoint's,
+    // matching every other DeepSeek V4 row in this catalog.
+    modelId: "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "vllm",
+    contextWindow: 1e6,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    modelId: "deepseek-ai/DeepSeek-V4-Pro",
+    displayName: "DeepSeek V4 Pro",
+    provider: "vllm",
+    contextWindow: 1e6,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.5-0.8B",
+    displayName: "Qwen 3.5 0.8B",
+    provider: "vllm",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.5-9B",
+    displayName: "Qwen 3.5 9B",
+    provider: "vllm",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.6-35B-A3B",
+    displayName: "Qwen 3.6 35B A3B",
+    provider: "vllm",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.8-27B",
+    displayName: "Qwen 3.8 27B",
+    provider: "vllm",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  {
+    modelId: "Qwen/Qwen3.8-Flash-Next",
+    displayName: "Qwen 3.8 Flash Next",
+    provider: "vllm",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: true,
+    clientType: VLLM_CLIENT_TYPE,
+    baseUrl: VLLM_LOCAL_BASE_URL
+  },
+  // -- Custom (the group that otherwise holds only user-defined models). A preset may live
+  // here only as a complete row — its own base URL and a pinned generic protocol — because the
+  // group implies neither; it is where a vendor with a single preview model and no console
+  // of its own goes rather than opening a group for it. Atria Dawn: the Anthropic
+  // Messages API at api.atria-asi.ai (the client appends /v1/messages, so the base URL carries
+  // no /v1; the endpoint also serves Chat Completions and Responses, but its Responses side
+  // rejects the replayed assistant turn of a multi-turn conversation), a 256K window (262144 —
+  // the API caps max_output_tokens at that minus the input), text only (the endpoint rejects
+  // image input), and no published price yet, so the row records $0 until the vendor prices
+  // it. Read 2026-09-15 from api.atria-asi.ai/docs.
+  {
+    modelId: "Atria-Dawn",
+    displayName: "Atria Dawn",
+    provider: "custom",
+    contextWindow: 262144,
+    pricing: usd(0, 0, 0),
+    supportsVision: false,
+    clientType: "ant-messages",
+    baseUrl: "https://api.atria-asi.ai"
+  }
+];
+function catalogEntryFor(provider, upstreamId) {
+  return MODEL_CATALOG.find((m) => m.provider === provider && m.modelId === upstreamId);
+}
+function providerInfo(providerId) {
+  return MODEL_PROVIDERS.find((p) => p.id === providerId);
+}
+function providerClientType(providerId) {
+  return providerInfo(providerId)?.clientType;
+}
+function isVendorGroup(providerId) {
+  const info = providerInfo(providerId);
+  return info !== void 0 && info.id !== "custom" && info.gatewayBaseUrl === void 0 && info.clientType === void 0;
+}
+function unroutableVendorModel(provider, modelId, clientType) {
+  const id = modelId.trim();
+  if (id === "" || !isVendorGroup(provider)) return false;
+  return routedClientType(id, clientType?.trim() ? clientType : null) === null;
+}
+function vendorClientType(providerId) {
+  if (!isVendorGroup(providerId)) return void 0;
+  const envKey = providerInfo(providerId)?.envKey;
+  return OFFICIAL_CLIENT_TYPES.find((t) => CLIENT_CREDENTIALS[t]?.apiKeyEnv === envKey);
+}
+function isAddableGroup(providerId) {
+  const info = providerInfo(providerId);
+  return info === void 0 || info.addable === true;
+}
+function unaddableModel(provider, modelId) {
+  return !isAddableGroup(provider) && catalogEntryFor(provider, modelId.trim()) === void 0;
+}
+function retainedCatalogPricing(provider, modelId) {
+  const entry = catalogEntryFor(provider, modelId);
+  return entry?.wasDefault === true ? entry.pricing : void 0;
+}
+function modelHomepageUrl(provider, modelId) {
+  if (provider === "openrouter") return `https://openrouter.ai/${modelId}`;
+  if (provider === "qwen-token-plan") {
+    return `https://www.qianwenai.com/models/${modelId}`;
+  }
+  if (provider === "fireworks") {
+    const m = /^accounts\/([^/]+)\/models\/(.+)$/.exec(modelId);
+    return m ? `https://app.fireworks.ai/models/${m[1]}/${m[2]}` : providerInfo(provider)?.modelsUrl;
+  }
+  if (provider === "tokendance") return `https://tokendance.space/models/${modelId}`;
+  if (provider === "qwen-pay-as-you-go") {
+    return `https://www.qianwenai.com/models/${encodeURIComponent(modelId)}`;
+  }
+  if (provider === "zhipu") {
+    return `https://docs.z.ai/guides/llm/${modelId}`;
+  }
+  if (provider === "moonshot") {
+    const m = /^kimi-k(\d+)\.(\d+)$/.exec(modelId);
+    return m ? `https://platform.kimi.com/docs/pricing/chat-k${m[1]}${m[2]}` : providerInfo(provider)?.modelsUrl;
+  }
+  if (provider === "vllm") {
+    return catalogEntryFor(provider, modelId) !== void 0 ? `https://recipes.vllm.ai/${modelId}` : providerInfo(provider)?.modelsUrl;
+  }
+  if (provider === "custom") return void 0;
+  return providerInfo(provider)?.modelsUrl;
+}
+
+// src/model-catalog/env-keys.ts
+import {
+  CLIENT_CREDENTIALS as CLIENT_CREDENTIALS2,
+  environmentCredential,
+  routedClientType as routedClientType2
+} from "@prismshadow/mmsp/credentials";
+import { effectiveConnection } from "@prismshadow/penguin-core/model-connection";
+function clientEnvKey(clientType, baseUrl) {
+  const credential = environmentCredential(clientType, baseUrl?.trim() || void 0);
+  return credential?.kind === "env" ? credential.apiKeyEnv : void 0;
+}
+function calledKeyless(clientType, baseUrl) {
+  const credential = environmentCredential(clientType, baseUrl?.trim() || void 0);
+  if (credential?.kind === "keyless") return true;
+  if (credential?.kind !== "env") return false;
+  const routed = routedClientType2("", clientType);
+  return routed !== null && CLIENT_CREDENTIALS2[routed]?.openServer === true;
+}
+function signInKeyed(providerId) {
+  return providerInfo(providerId)?.bridgeAuth !== void 0;
+}
+function modelEnvironment(entry, group) {
+  const effective = effectiveConnection(entry, group);
+  const routed = routedClientType2(entry.modelId, effective.clientType ?? null);
+  if (routed === null) return {};
+  const credential = environmentCredential(routed, effective.baseUrl);
+  if (credential?.kind !== "env") return {};
+  const defaultBaseUrl = CLIENT_CREDENTIALS2[routed]?.defaultBaseUrl;
+  return {
+    ...effective.apiKey === void 0 ? { apiKeyEnv: credential.apiKeyEnv } : {},
+    ...credential.baseUrlEnv !== null ? {
+      baseUrlEnv: credential.baseUrlEnv,
+      ...defaultBaseUrl !== void 0 ? { defaultBaseUrl } : {}
+    } : {}
+  };
+}
+function modelEnvPreviewKey(entry, group) {
+  return modelEnvironment(entry, group).apiKeyEnv;
+}
+function groupClientType(providerId, group) {
+  const pinned = group?.clientType?.trim() ? group.clientType : void 0;
+  return pinned ?? vendorClientType(providerId);
+}
+function providerEnvFallbackKey(providerId, group) {
+  const clientType = groupClientType(providerId, group);
+  return clientType === void 0 ? void 0 : clientEnvKey(clientType, group?.baseUrl);
+}
+function providerCalledKeyless(providerId, group) {
+  if (signInKeyed(providerId)) return false;
+  const clientType = groupClientType(providerId, group);
+  return clientType !== void 0 && calledKeyless(clientType, group?.baseUrl);
+}
+function modelCalledKeyless(entry, group) {
+  if (signInKeyed(entry.provider)) return false;
+  const effective = effectiveConnection(entry, group);
+  if (effective.apiKey !== void 0) return false;
+  const routed = routedClientType2(entry.modelId, effective.clientType ?? null);
+  return routed !== null && calledKeyless(routed, effective.baseUrl);
+}
+
+// src/model-catalog/seed.ts
+import {
+  canonicalClientType,
+  sameClientType as sameClientType2,
+  sameEndpoint
+} from "@prismshadow/penguin-core/model-connection";
+function sharedValue(values, same) {
+  const first = values[0];
+  if (first === void 0) return void 0;
+  return values.every((v) => v !== void 0 && same(v, first)) ? first : void 0;
+}
+function catalogGroupConnection(providerId) {
+  const info = providerInfo(providerId);
+  if (info === void 0 || info.id === "custom") return void 0;
+  const rows = MODEL_CATALOG.filter((m) => m.provider === providerId);
+  const baseUrl = info.gatewayBaseUrl ?? sharedValue(
+    rows.map((m) => m.baseUrl),
+    sameEndpoint
+  );
+  const clientType = canonicalClientType(
+    providerClientType(providerId) ?? sharedValue(
+      rows.map((m) => m.clientType),
+      sameClientType2
+    )
+  );
+  if (baseUrl === void 0 && clientType === void 0) return void 0;
+  return {
+    ...baseUrl !== void 0 ? { base_url: baseUrl } : {},
+    ...clientType !== void 0 ? { client_type: clientType } : {}
+  };
+}
+var PRESET_DEFAULT_MODEL = {
+  provider: "deepseek",
+  model_id: "deepseek-flash"
+};
+function presetProviderTable() {
+  const table = {};
+  for (const p of MODEL_PROVIDERS) {
+    const connection = catalogGroupConnection(p.id);
+    if (connection !== void 0) table[p.id] = connection;
+  }
+  return table;
+}
+function presetModelEntries(groups = presetProviderTable()) {
+  const catalog = presetProviderTable();
+  return MODEL_CATALOG.filter((m) => m.retired !== true).map(
+    (m) => catalogModelEntry(m, catalog, groups)
+  );
+}
+function catalogModelEntries(groups = presetProviderTable()) {
+  const catalog = presetProviderTable();
+  return MODEL_CATALOG.map((m) => catalogModelEntry(m, catalog, groups));
+}
+function catalogModelEntry(m, catalog, groups) {
+  const group = groups[m.provider];
+  const clientType = canonicalClientType(m.clientType ?? catalog[m.provider]?.client_type);
+  const groupClientType2 = group?.client_type?.trim() || void 0;
+  const ownClientType = clientType !== void 0 && (groupClientType2 === void 0 || !sameClientType2(clientType, groupClientType2)) ? clientType : void 0;
+  const baseUrl = m.baseUrl ?? catalog[m.provider]?.base_url;
+  const groupBaseUrl = group?.base_url?.trim() || void 0;
+  const ownBaseUrl = baseUrl !== void 0 && (groupBaseUrl === void 0 || !sameEndpoint(baseUrl, groupBaseUrl)) ? baseUrl : void 0;
+  return {
+    provider: m.provider,
+    model_id: m.modelId,
+    ...m.contextWindow !== void 0 ? { context_window: m.contextWindow } : {},
+    ...ownClientType !== void 0 ? { client_type: ownClientType } : {},
+    ...m.pricing ? { pricing: { ...m.pricing } } : {},
+    // ModelEntry.vision defaults to supported: only models that don't support images
+    // explicitly persist false (drives read_file's hand-off of images to the vision model and input
+    // image hand-off, see project-config.ts).
+    ...m.supportsVision ? {} : { vision: false },
+    ...ownBaseUrl !== void 0 ? { base_url: ownBaseUrl } : {}
+  };
+}
+function presetPromotions() {
+  return MODEL_CATALOG.flatMap(
+    (m) => m.retired !== true && m.discount !== void 0 && m.discount > 0 && m.discount < 1 ? [{ provider: m.provider, modelId: m.modelId, discount: m.discount }] : []
+  );
+}
+
+export {
+  PENGUIN_GO_BASE_URL,
+  PENGUIN_GO_PROVIDER_ID,
+  fixedClientType,
+  fixedClientTypeRefuses,
+  MODELSCOPE_PROVIDER_ID,
+  MODEL_PROVIDERS,
+  offPeakAt,
+  DEEPSEEK_OFF_PEAK,
+  QWEN_OFF_PEAK,
+  offPeakScheduledRefs,
+  effectivePricing,
+  MODEL_CATALOG,
+  catalogEntryFor,
+  providerInfo,
+  providerClientType,
+  isVendorGroup,
+  unroutableVendorModel,
+  vendorClientType,
+  isAddableGroup,
+  unaddableModel,
+  retainedCatalogPricing,
+  modelHomepageUrl,
+  clientEnvKey,
+  modelEnvironment,
+  modelEnvPreviewKey,
+  providerEnvFallbackKey,
+  providerCalledKeyless,
+  modelCalledKeyless,
+  catalogGroupConnection,
+  PRESET_DEFAULT_MODEL,
+  presetProviderTable,
+  presetModelEntries,
+  catalogModelEntries,
+  presetPromotions
+};
+//# sourceMappingURL=chunk-GGIINEEK.js.map