agent-duelist 0.2.1 → 0.3.1

package/dist/cli.js

@@ -498,7 +498,7 @@ function Pe(e5, t3, s5 = Q.DEFAULT) {
     return p5(), n3.getToken() !== 2 ? k3(7, [2], []) : b3(), true;
   }
   i2(M2, "parseObject");
-  function z2() {
+  function z5() {
     D3(), b3();
     let w4 = true, j2 = false;
     for (; n3.getToken() !== 4 && n3.getToken() !== 17; ) {
@@ -509,11 +509,11 @@ function Pe(e5, t3, s5 = Q.DEFAULT) {
     }
     return L3(), w4 || o7.pop(), n3.getToken() !== 4 ? k3(8, [4], []) : b3(), true;
   }
-  i2(z2, "parseArray");
+  i2(z5, "parseArray");
   function U2() {
     switch (n3.getToken()) {
       case 3:
-        return z2();
+        return z5();
       case 1:
         return M2();
       case 10:
@@ -746,7 +746,7 @@ var init_dist2 = __esm({
     }, "interpolateConfigDir");
     Me = ["outDir", "declarationDir", "outFile", "rootDir", "baseUrl", "tsBuildInfoFile"];
     ze = i2((e5) => {
-      var t3, s5, n3, o7, l3, u5, a7, r3, g2, v4, d5, _4, p5, D3, L3, T3, F4, x, c3, y3, A3, b3, k3, R5, W, V2, M2, z2, U2, w4, j2, S2, $2;
+      var t3, s5, n3, o7, l3, u5, a7, r3, g2, v4, d5, _4, p5, D3, L3, T3, F4, x, c3, y3, A3, b3, k3, R5, W, V2, M2, z5, U2, w4, j2, S2, $2;
       if (e5.strict) {
         const f6 = ["noImplicitAny", "noImplicitThis", "strictNullChecks", "strictFunctionTypes", "strictBindCallApply", "strictPropertyInitialization", "strictBuiltinIteratorReturn", "alwaysStrict", "useUnknownInCatchVariables"];
         for (const B2 of f6) e5[B2] === void 0 && (e5[B2] = true);
@@ -767,7 +767,7 @@ var init_dist2 = __esm({
         let f6 = e5.moduleResolution.toLowerCase();
         f6 === "node" && (f6 = "node10"), e5.moduleResolution = f6, (f6 === "node16" || f6 === "nodenext" || f6 === "bundler") && ((R5 = e5.resolvePackageJsonExports) != null || (e5.resolvePackageJsonExports = true), (W = e5.resolvePackageJsonImports) != null || (e5.resolvePackageJsonImports = true)), f6 === "bundler" && ((V2 = e5.allowSyntheticDefaultImports) != null || (e5.allowSyntheticDefaultImports = true), (M2 = e5.resolveJsonModule) != null || (e5.resolveJsonModule = true));
       }
-      e5.jsx && (e5.jsx = e5.jsx.toLowerCase()), e5.moduleDetection && (e5.moduleDetection = e5.moduleDetection.toLowerCase()), e5.importsNotUsedAsValues && (e5.importsNotUsedAsValues = e5.importsNotUsedAsValues.toLowerCase()), e5.newLine && (e5.newLine = e5.newLine.toLowerCase()), e5.esModuleInterop && ((z2 = e5.allowSyntheticDefaultImports) != null || (e5.allowSyntheticDefaultImports = true)), e5.verbatimModuleSyntax && ((U2 = e5.isolatedModules) != null || (e5.isolatedModules = true), (w4 = e5.preserveConstEnums) != null || (e5.preserveConstEnums = true)), e5.isolatedModules && ((j2 = e5.preserveConstEnums) != null || (e5.preserveConstEnums = true)), e5.rewriteRelativeImportExtensions && ((S2 = e5.allowImportingTsExtensions) != null || (e5.allowImportingTsExtensions = true)), e5.lib && (e5.lib = e5.lib.map((f6) => f6.toLowerCase())), e5.checkJs && (($2 = e5.allowJs) != null || (e5.allowJs = true));
+      e5.jsx && (e5.jsx = e5.jsx.toLowerCase()), e5.moduleDetection && (e5.moduleDetection = e5.moduleDetection.toLowerCase()), e5.importsNotUsedAsValues && (e5.importsNotUsedAsValues = e5.importsNotUsedAsValues.toLowerCase()), e5.newLine && (e5.newLine = e5.newLine.toLowerCase()), e5.esModuleInterop && ((z5 = e5.allowSyntheticDefaultImports) != null || (e5.allowSyntheticDefaultImports = true)), e5.verbatimModuleSyntax && ((U2 = e5.isolatedModules) != null || (e5.isolatedModules = true), (w4 = e5.preserveConstEnums) != null || (e5.preserveConstEnums = true)), e5.isolatedModules && ((j2 = e5.preserveConstEnums) != null || (e5.preserveConstEnums = true)), e5.rewriteRelativeImportExtensions && ((S2 = e5.allowImportingTsExtensions) != null || (e5.allowImportingTsExtensions = true)), e5.lib && (e5.lib = e5.lib.map((f6) => f6.toLowerCase())), e5.checkJs && (($2 = e5.allowJs) != null || (e5.allowJs = true));
     }, "normalizeCompilerOptions");
     pe = i2((e5, t3 = /* @__PURE__ */ new Map()) => {
       const s5 = m3.resolve(e5), n3 = ve(s5, t3), o7 = m3.dirname(s5), { compilerOptions: l3 } = n3;
@@ -3345,14 +3345,14 @@ function fn(s5, e5 = "@") {
               case 32:
                 break;
               case 101: {
-                if (!(d5[400] | 0) && z2(h4) | 0 && !(A3(t3 + 4 | 0, 16, 10) | 0) && ($2(), (b3[804] | 0) == 0)) {
+                if (!(d5[400] | 0) && z5(h4) | 0 && !(A3(t3 + 4 | 0, 16, 10) | 0) && ($2(), (b3[804] | 0) == 0)) {
                   l3 = 9;
                   break e;
                 } else l3 = 17;
                 break;
               }
               case 105: {
-                z2(h4) | 0 && !(A3(t3 + 4 | 0, 26, 10) | 0) && W(), l3 = 17;
+                z5(h4) | 0 && !(A3(t3 + 4 | 0, 26, 10) | 0) && W(), l3 = 17;
                 break;
               }
               case 59: {
@@ -3400,15 +3400,15 @@ function fn(s5, e5 = "@") {
                   case 32:
                     break;
                   case 101: {
-                    !(d5[400] | 0) && z2(a7) | 0 && !(A3(t3 + 4 | 0, 16, 10) | 0) && $2(), l3 = 91;
+                    !(d5[400] | 0) && z5(a7) | 0 && !(A3(t3 + 4 | 0, 16, 10) | 0) && $2(), l3 = 91;
                     break;
                   }
                   case 105: {
-                    z2(a7) | 0 && !(A3(t3 + 4 | 0, 26, 10) | 0) && W(), l3 = 91;
+                    z5(a7) | 0 && !(A3(t3 + 4 | 0, 26, 10) | 0) && W(), l3 = 91;
                     break;
                   }
                   case 99: {
-                    z2(a7) | 0 && !(A3(t3 + 4 | 0, 36, 8) | 0) && P3(d5[t3 + 12 >> 1] | 0) | 0 && (b3[806] = 1), l3 = 91;
+                    z5(a7) | 0 && !(A3(t3 + 4 | 0, 36, 8) | 0) && P3(d5[t3 + 12 >> 1] | 0) | 0 && (b3[806] = 1), l3 = 91;
                     break;
                   }
                   case 40: {
@@ -4437,10 +4437,10 @@ function fn(s5, e5 = "@") {
         return t3 = t3 | 0, (d5[t3 >> 1] | 0) == 46 && (d5[t3 + -2 >> 1] | 0) == 46 ? t3 = (d5[t3 + -4 >> 1] | 0) == 46 : t3 = 0, t3 | 0;
       }
       f2(de3, "G");
-      function z2(t3) {
+      function z5(t3) {
         return t3 = t3 | 0, (r3[3] | 0) == (t3 | 0) ? t3 = 1 : t3 = Oe2(t3 + -2 | 0) | 0, t3 | 0;
       }
-      f2(z2, "H");
+      f2(z5, "H");
       function vt() {
         var t3 = 0;
         return t3 = r3[(r3[62] | 0) + 12 >> 2] | 0, t3 ? t3 = t3 - (r3[3] | 0) >> 1 : t3 = -1, t3 | 0;
@@ -5821,7 +5821,7 @@ import p4 from "path";
 import { fileURLToPath as O4 } from "url";
 import se3, { writeSync as te2 } from "fs";
 import { inspect as oe3 } from "util";
-var K3, o4, R4, D2, me3, N2, j, pe2, y2, C3, de2, E4, ge2, Q4, M, _3, S, A2, T2, Pe3, I4, F3, v3, J3, P2, je3, be2, xe3, k2, $, ye3, Ee, B, G3, _e3, Se3, b2, X3, w2, ve3, z, we2, Me3, Te3, Fe3, H2, $e3;
+var K3, o4, R4, D2, me3, N2, j, pe2, y2, C3, de2, E4, ge2, Q4, M, _3, S, A2, T2, Pe3, I4, F3, v3, J3, P2, je3, be2, xe3, k2, $, ye3, Ee, B, G3, _e3, Se3, b2, X3, w2, ve3, z4, we2, Me3, Te3, Fe3, H2, $e3;
 var init_register_CFH5oNdT = __esm({
   "node_modules/tsx/dist/register-CFH5oNdT.mjs"() {
     "use strict";
@@ -5995,11 +5995,11 @@ var init_register_CFH5oNdT = __esm({
         throw t3;
       }
     }, "createTsExtensionResolver");
-    z = "at cjsPreparseModuleExports (node:internal";
+    z4 = "at cjsPreparseModuleExports (node:internal";
     we2 = o4((s5) => {
       const e5 = s5.stack.split(`
 `).slice(1);
-      return e5[1].includes(z) || e5[2].includes(z);
+      return e5[1].includes(z4) || e5[2].includes(z4);
     }, "isFromCjsLexer");
     Me3 = o4((s5, e5) => {
       const a7 = s5.split("?"), n3 = new URLSearchParams(a7[1]);
@@ -6197,6 +6197,1748 @@ import { readFileSync as readFileSync3, writeFileSync as writeFileSync2, mkdirSy
 import { resolve, join, dirname as dirname2 } from "path";
 import { pathToFileURL, fileURLToPath } from "url";
 
+// src/scorers/latency.ts
+var MIN_MS = 500;
+var MAX_MS = 1e4;
+var latencyScorer = ({ result }) => {
+  const clamped = Math.max(MIN_MS, Math.min(MAX_MS, result.latencyMs));
+  const value = 1 - (clamped - MIN_MS) / (MAX_MS - MIN_MS);
+  return {
+    name: "latency",
+    value: Math.round(value * 100) / 100,
+    details: { ms: result.latencyMs }
+  };
+};
+
+// src/pricing/catalog.json
+var catalog_default = {
+  _meta: {
+    source: "OpenRouter API \u2014 all providers (https://openrouter.ai/api/v1/models)",
+    updatedAt: "2026-02-28",
+    unit: "USD per token"
+  },
+  models: {
+    "ai21/jamba-large-1.7": {
+      inputPerToken: 2e-6,
+      outputPerToken: 8e-6
+    },
+    "aion-labs/aion-1.0": {
+      inputPerToken: 4e-6,
+      outputPerToken: 8e-6
+    },
+    "aion-labs/aion-1.0-mini": {
+      inputPerToken: 7e-7,
+      outputPerToken: 14e-7
+    },
+    "aion-labs/aion-2.0": {
+      inputPerToken: 8e-7,
+      outputPerToken: 16e-7
+    },
+    "aion-labs/aion-rp-llama-3.1-8b": {
+      inputPerToken: 8e-7,
+      outputPerToken: 16e-7
+    },
+    "alfredpros/codellama-7b-instruct-solidity": {
+      inputPerToken: 8e-7,
+      outputPerToken: 12e-7
+    },
+    "alibaba/tongyi-deepresearch-30b-a3b": {
+      inputPerToken: 9e-8,
+      outputPerToken: 45e-8
+    },
+    "allenai/molmo-2-8b": {
+      inputPerToken: 2e-7,
+      outputPerToken: 2e-7
+    },
+    "allenai/olmo-2-0325-32b-instruct": {
+      inputPerToken: 5e-8,
+      outputPerToken: 2e-7
+    },
+    "allenai/olmo-3-32b-think": {
+      inputPerToken: 15e-8,
+      outputPerToken: 5e-7
+    },
+    "allenai/olmo-3-7b-instruct": {
+      inputPerToken: 1e-7,
+      outputPerToken: 2e-7
+    },
+    "allenai/olmo-3-7b-think": {
+      inputPerToken: 12e-8,
+      outputPerToken: 2e-7
+    },
+    "allenai/olmo-3.1-32b-instruct": {
+      inputPerToken: 2e-7,
+      outputPerToken: 6e-7
+    },
+    "alpindale/goliath-120b": {
+      inputPerToken: 375e-8,
+      outputPerToken: 75e-7
+    },
+    "amazon/nova-2-lite-v1": {
+      inputPerToken: 3e-7,
+      outputPerToken: 25e-7
+    },
+    "amazon/nova-lite-v1": {
+      inputPerToken: 6e-8,
+      outputPerToken: 24e-8
+    },
+    "amazon/nova-micro-v1": {
+      inputPerToken: 35e-9,
+      outputPerToken: 14e-8
+    },
+    "amazon/nova-premier-v1": {
+      inputPerToken: 25e-7,
+      outputPerToken: 125e-7
+    },
+    "amazon/nova-pro-v1": {
+      inputPerToken: 8e-7,
+      outputPerToken: 32e-7
+    },
+    "anthracite-org/magnum-v4-72b": {
+      inputPerToken: 3e-6,
+      outputPerToken: 5e-6
+    },
+    "anthropic/claude-3-haiku": {
+      inputPerToken: 25e-8,
+      outputPerToken: 125e-8
+    },
+    "anthropic/claude-3.5-haiku": {
+      inputPerToken: 8e-7,
+      outputPerToken: 4e-6
+    },
+    "anthropic/claude-3.5-sonnet": {
+      inputPerToken: 6e-6,
+      outputPerToken: 3e-5
+    },
+    "anthropic/claude-3.7-sonnet": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "anthropic/claude-3.7-sonnet:thinking": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "anthropic/claude-haiku-4.5": {
+      inputPerToken: 1e-6,
+      outputPerToken: 5e-6
+    },
+    "anthropic/claude-opus-4": {
+      inputPerToken: 15e-6,
+      outputPerToken: 75e-6
+    },
+    "anthropic/claude-opus-4.1": {
+      inputPerToken: 15e-6,
+      outputPerToken: 75e-6
+    },
+    "anthropic/claude-opus-4.5": {
+      inputPerToken: 5e-6,
+      outputPerToken: 25e-6
+    },
+    "anthropic/claude-opus-4.6": {
+      inputPerToken: 5e-6,
+      outputPerToken: 25e-6
+    },
+    "anthropic/claude-sonnet-4": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "anthropic/claude-sonnet-4.5": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "anthropic/claude-sonnet-4.6": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "arcee-ai/coder-large": {
+      inputPerToken: 5e-7,
+      outputPerToken: 8e-7
+    },
+    "arcee-ai/maestro-reasoning": {
+      inputPerToken: 9e-7,
+      outputPerToken: 33e-7
+    },
+    "arcee-ai/spotlight": {
+      inputPerToken: 18e-8,
+      outputPerToken: 18e-8
+    },
+    "arcee-ai/trinity-mini": {
+      inputPerToken: 45e-9,
+      outputPerToken: 15e-8
+    },
+    "arcee-ai/virtuoso-large": {
+      inputPerToken: 75e-8,
+      outputPerToken: 12e-7
+    },
+    "baidu/ernie-4.5-21b-a3b": {
+      inputPerToken: 7e-8,
+      outputPerToken: 28e-8
+    },
+    "baidu/ernie-4.5-21b-a3b-thinking": {
+      inputPerToken: 7e-8,
+      outputPerToken: 28e-8
+    },
+    "baidu/ernie-4.5-300b-a47b": {
+      inputPerToken: 28e-8,
+      outputPerToken: 11e-7
+    },
+    "baidu/ernie-4.5-vl-28b-a3b": {
+      inputPerToken: 14e-8,
+      outputPerToken: 56e-8
+    },
+    "baidu/ernie-4.5-vl-424b-a47b": {
+      inputPerToken: 42e-8,
+      outputPerToken: 125e-8
+    },
+    "bytedance/seed-1.6": {
+      inputPerToken: 25e-8,
+      outputPerToken: 2e-6
+    },
+    "bytedance/seed-1.6-flash": {
+      inputPerToken: 75e-9,
+      outputPerToken: 3e-7
+    },
+    "bytedance/seed-2.0-mini": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "bytedance/ui-tars-1.5-7b": {
+      inputPerToken: 1e-7,
+      outputPerToken: 2e-7
+    },
+    "cohere/command-a": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "cohere/command-r-08-2024": {
+      inputPerToken: 15e-8,
+      outputPerToken: 6e-7
+    },
+    "cohere/command-r-plus-08-2024": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "cohere/command-r7b-12-2024": {
+      inputPerToken: 375e-10,
+      outputPerToken: 15e-8
+    },
+    "deepcogito/cogito-v2.1-671b": {
+      inputPerToken: 125e-8,
+      outputPerToken: 125e-8
+    },
+    "deepseek/deepseek-chat": {
+      inputPerToken: 32e-8,
+      outputPerToken: 89e-8
+    },
+    "deepseek/deepseek-chat-v3-0324": {
+      inputPerToken: 2e-7,
+      outputPerToken: 77e-8
+    },
+    "deepseek/deepseek-chat-v3.1": {
+      inputPerToken: 15e-8,
+      outputPerToken: 75e-8
+    },
+    "deepseek/deepseek-r1": {
+      inputPerToken: 7e-7,
+      outputPerToken: 25e-7
+    },
+    "deepseek/deepseek-r1-0528": {
+      inputPerToken: 45e-8,
+      outputPerToken: 215e-8
+    },
+    "deepseek/deepseek-r1-distill-llama-70b": {
+      inputPerToken: 7e-7,
+      outputPerToken: 8e-7
+    },
+    "deepseek/deepseek-r1-distill-qwen-32b": {
+      inputPerToken: 29e-8,
+      outputPerToken: 29e-8
+    },
+    "deepseek/deepseek-v3": {
+      inputPerToken: 3e-7,
+      outputPerToken: 88e-8
+    },
+    "deepseek/deepseek-v3.1-terminus": {
+      inputPerToken: 21e-8,
+      outputPerToken: 79e-8
+    },
+    "deepseek/deepseek-v3.1-terminus:exacto": {
+      inputPerToken: 21e-8,
+      outputPerToken: 79e-8
+    },
+    "deepseek/deepseek-v3.2": {
+      inputPerToken: 25e-8,
+      outputPerToken: 4e-7
+    },
+    "deepseek/deepseek-v3.2-exp": {
+      inputPerToken: 27e-8,
+      outputPerToken: 41e-8
+    },
+    "deepseek/deepseek-v3.2-speciale": {
+      inputPerToken: 4e-7,
+      outputPerToken: 12e-7
+    },
+    "eleutherai/llemma_7b": {
+      inputPerToken: 8e-7,
+      outputPerToken: 12e-7
+    },
+    "essentialai/rnj-1-instruct": {
+      inputPerToken: 15e-8,
+      outputPerToken: 15e-8
+    },
+    "google/gemini-2.0-flash": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "google/gemini-2.0-flash-001": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "google/gemini-2.0-flash-lite-001": {
+      inputPerToken: 75e-9,
+      outputPerToken: 3e-7
+    },
+    "google/gemini-2.5-flash": {
+      inputPerToken: 3e-7,
+      outputPerToken: 25e-7
+    },
+    "google/gemini-2.5-flash-image": {
+      inputPerToken: 3e-7,
+      outputPerToken: 25e-7
+    },
+    "google/gemini-2.5-flash-lite": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "google/gemini-2.5-flash-lite-preview-09-2025": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "google/gemini-2.5-pro": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "google/gemini-2.5-pro-preview": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "google/gemini-2.5-pro-preview-05-06": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "google/gemini-3-flash-preview": {
+      inputPerToken: 5e-7,
+      outputPerToken: 3e-6
+    },
+    "google/gemini-3-pro-image-preview": {
+      inputPerToken: 2e-6,
+      outputPerToken: 12e-6
+    },
+    "google/gemini-3-pro-preview": {
+      inputPerToken: 2e-6,
+      outputPerToken: 12e-6
+    },
+    "google/gemini-3.1-flash-image-preview": {
+      inputPerToken: 25e-8,
+      outputPerToken: 15e-7
+    },
+    "google/gemini-3.1-pro-preview": {
+      inputPerToken: 2e-6,
+      outputPerToken: 12e-6
+    },
+    "google/gemini-3.1-pro-preview-customtools": {
+      inputPerToken: 2e-6,
+      outputPerToken: 12e-6
+    },
+    "google/gemma-2-27b-it": {
+      inputPerToken: 65e-8,
+      outputPerToken: 65e-8
+    },
+    "google/gemma-2-9b-it": {
+      inputPerToken: 3e-8,
+      outputPerToken: 9e-8
+    },
+    "google/gemma-3-12b-it": {
+      inputPerToken: 4e-8,
+      outputPerToken: 13e-8
+    },
+    "google/gemma-3-27b-it": {
+      inputPerToken: 4e-8,
+      outputPerToken: 15e-8
+    },
+    "google/gemma-3-4b-it": {
+      inputPerToken: 4e-8,
+      outputPerToken: 8e-8
+    },
+    "google/gemma-3n-e4b-it": {
+      inputPerToken: 2e-8,
+      outputPerToken: 4e-8
+    },
+    "gryphe/mythomax-l2-13b": {
+      inputPerToken: 6e-8,
+      outputPerToken: 6e-8
+    },
+    "ibm-granite/granite-4.0-h-micro": {
+      inputPerToken: 17e-9,
+      outputPerToken: 11e-8
+    },
+    "inception/mercury": {
+      inputPerToken: 25e-8,
+      outputPerToken: 1e-6
+    },
+    "inception/mercury-coder": {
+      inputPerToken: 25e-8,
+      outputPerToken: 1e-6
+    },
+    "inflection/inflection-3-pi": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "inflection/inflection-3-productivity": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "kwaipilot/kat-coder-pro": {
+      inputPerToken: 207e-9,
+      outputPerToken: 828e-9
+    },
+    "liquid/lfm-2-24b-a2b": {
+      inputPerToken: 3e-8,
+      outputPerToken: 12e-8
+    },
+    "liquid/lfm-2.2-6b": {
+      inputPerToken: 1e-8,
+      outputPerToken: 2e-8
+    },
+    "liquid/lfm2-8b-a1b": {
+      inputPerToken: 1e-8,
+      outputPerToken: 2e-8
+    },
+    "mancer/weaver": {
+      inputPerToken: 75e-8,
+      outputPerToken: 1e-6
+    },
+    "meituan/longcat-flash-chat": {
+      inputPerToken: 2e-7,
+      outputPerToken: 8e-7
+    },
+    "meta/llama-3-70b-instruct": {
+      inputPerToken: 51e-8,
+      outputPerToken: 74e-8
+    },
+    "meta/llama-3-8b-instruct": {
+      inputPerToken: 3e-8,
+      outputPerToken: 4e-8
+    },
+    "meta/llama-3.1-405b": {
+      inputPerToken: 4e-6,
+      outputPerToken: 4e-6
+    },
+    "meta/llama-3.1-405b-instruct": {
+      inputPerToken: 4e-6,
+      outputPerToken: 4e-6
+    },
+    "meta/llama-3.1-70b-instruct": {
+      inputPerToken: 4e-7,
+      outputPerToken: 4e-7
+    },
+    "meta/llama-3.1-8b-instruct": {
+      inputPerToken: 2e-8,
+      outputPerToken: 5e-8
+    },
+    "meta/llama-3.2-11b-vision-instruct": {
+      inputPerToken: 49e-9,
+      outputPerToken: 49e-9
+    },
+    "meta/llama-3.2-1b-instruct": {
+      inputPerToken: 27e-9,
+      outputPerToken: 2e-7
+    },
+    "meta/llama-3.2-3b-instruct": {
+      inputPerToken: 2e-8,
+      outputPerToken: 2e-8
+    },
+    "meta/llama-3.3-70b": {
+      inputPerToken: 12e-8,
+      outputPerToken: 3e-7
+    },
+    "meta/llama-3.3-70b-instruct": {
+      inputPerToken: 1e-7,
+      outputPerToken: 32e-8
+    },
+    "meta/llama-4-maverick": {
+      inputPerToken: 15e-8,
+      outputPerToken: 6e-7
+    },
+    "meta/llama-4-scout": {
+      inputPerToken: 8e-8,
+      outputPerToken: 3e-7
+    },
+    "meta/llama-guard-2-8b": {
+      inputPerToken: 2e-7,
+      outputPerToken: 2e-7
+    },
+    "meta/llama-guard-3-8b": {
+      inputPerToken: 2e-8,
+      outputPerToken: 6e-8
+    },
+    "meta/llama-guard-4-12b": {
+      inputPerToken: 18e-8,
+      outputPerToken: 18e-8
+    },
+    "microsoft/phi-4": {
+      inputPerToken: 6e-8,
+      outputPerToken: 14e-8
+    },
+    "microsoft/wizardlm-2-8x22b": {
+      inputPerToken: 62e-8,
+      outputPerToken: 62e-8
+    },
+    "minimax/minimax-01": {
+      inputPerToken: 2e-7,
+      outputPerToken: 11e-7
+    },
+    "minimax/minimax-m1": {
+      inputPerToken: 4e-7,
+      outputPerToken: 22e-7
+    },
+    "minimax/minimax-m2": {
+      inputPerToken: 255e-9,
+      outputPerToken: 1e-6
+    },
+    "minimax/minimax-m2-her": {
+      inputPerToken: 3e-7,
+      outputPerToken: 12e-7
+    },
+    "minimax/minimax-m2.1": {
+      inputPerToken: 27e-8,
+      outputPerToken: 95e-8
+    },
+    "minimax/minimax-m2.5": {
+      inputPerToken: 295e-9,
+      outputPerToken: 12e-7
+    },
+    "mistral/codestral-2508": {
+      inputPerToken: 3e-7,
+      outputPerToken: 9e-7
+    },
+    "mistral/devstral-2512": {
+      inputPerToken: 4e-7,
+      outputPerToken: 2e-6
+    },
+    "mistral/devstral-medium": {
+      inputPerToken: 4e-7,
+      outputPerToken: 2e-6
+    },
+    "mistral/devstral-small": {
+      inputPerToken: 1e-7,
+      outputPerToken: 3e-7
+    },
+    "mistral/ministral-14b-2512": {
+      inputPerToken: 2e-7,
+      outputPerToken: 2e-7
+    },
+    "mistral/ministral-3b-2512": {
+      inputPerToken: 1e-7,
+      outputPerToken: 1e-7
+    },
+    "mistral/ministral-8b-2512": {
+      inputPerToken: 15e-8,
+      outputPerToken: 15e-8
+    },
+    "mistral/mistral-7b-instruct": {
+      inputPerToken: 2e-7,
+      outputPerToken: 2e-7
+    },
+    "mistral/mistral-7b-instruct-v0.1": {
+      inputPerToken: 11e-8,
+      outputPerToken: 19e-8
+    },
+    "mistral/mistral-7b-instruct-v0.3": {
+      inputPerToken: 2e-7,
+      outputPerToken: 2e-7
+    },
+    "mistral/mistral-large": {
+      inputPerToken: 2e-6,
+      outputPerToken: 6e-6
+    },
+    "mistral/mistral-large-2407": {
+      inputPerToken: 2e-6,
+      outputPerToken: 6e-6
+    },
+    "mistral/mistral-large-2411": {
+      inputPerToken: 2e-6,
+      outputPerToken: 6e-6
+    },
+    "mistral/mistral-large-2512": {
+      inputPerToken: 5e-7,
+      outputPerToken: 15e-7
+    },
+    "mistral/mistral-medium-3": {
+      inputPerToken: 4e-7,
+      outputPerToken: 2e-6
+    },
+    "mistral/mistral-medium-3.1": {
+      inputPerToken: 4e-7,
+      outputPerToken: 2e-6
+    },
+    "mistral/mistral-nemo": {
+      inputPerToken: 2e-8,
+      outputPerToken: 4e-8
+    },
+    "mistral/mistral-saba": {
+      inputPerToken: 2e-7,
+      outputPerToken: 6e-7
+    },
+    "mistral/mistral-small": {
+      inputPerToken: 1e-7,
+      outputPerToken: 3e-7
+    },
+    "mistral/mistral-small-24b-instruct-2501": {
+      inputPerToken: 5e-8,
+      outputPerToken: 8e-8
+    },
+    "mistral/mistral-small-3.1-24b-instruct": {
+      inputPerToken: 35e-8,
+      outputPerToken: 56e-8
+    },
+    "mistral/mistral-small-3.2-24b-instruct": {
+      inputPerToken: 6e-8,
+      outputPerToken: 18e-8
+    },
+    "mistral/mistral-small-creative": {
+      inputPerToken: 1e-7,
+      outputPerToken: 3e-7
+    },
+    "mistral/mixtral-8x22b-instruct": {
+      inputPerToken: 2e-6,
+      outputPerToken: 6e-6
+    },
+    "mistral/mixtral-8x7b-instruct": {
+      inputPerToken: 54e-8,
+      outputPerToken: 54e-8
+    },
+    "mistral/pixtral-large-2411": {
+      inputPerToken: 2e-6,
+      outputPerToken: 6e-6
+    },
+    "mistral/voxtral-small-24b-2507": {
+      inputPerToken: 1e-7,
+      outputPerToken: 3e-7
+    },
+    "moonshotai/kimi-k2": {
+      inputPerToken: 55e-8,
+      outputPerToken: 22e-7
+    },
+    "moonshotai/kimi-k2-0905": {
+      inputPerToken: 4e-7,
+      outputPerToken: 2e-6
+    },
+    "moonshotai/kimi-k2-0905:exacto": {
+      inputPerToken: 6e-7,
+      outputPerToken: 25e-7
+    },
+    "moonshotai/kimi-k2-thinking": {
+      inputPerToken: 47e-8,
+      outputPerToken: 2e-6
+    },
+    "moonshotai/kimi-k2.5": {
+      inputPerToken: 45e-8,
+      outputPerToken: 22e-7
+    },
+    "morph/morph-v3-fast": {
+      inputPerToken: 8e-7,
+      outputPerToken: 12e-7
+    },
+    "morph/morph-v3-large": {
+      inputPerToken: 9e-7,
+      outputPerToken: 19e-7
+    },
+    "neversleep/llama-3.1-lumimaid-8b": {
+      inputPerToken: 9e-8,
+      outputPerToken: 6e-7
+    },
+    "neversleep/noromaid-20b": {
+      inputPerToken: 1e-6,
+      outputPerToken: 175e-8
+    },
+    "nex-agi/deepseek-v3.1-nex-n1": {
+      inputPerToken: 27e-8,
+      outputPerToken: 1e-6
+    },
+    "nousresearch/hermes-2-pro-llama-3-8b": {
+      inputPerToken: 14e-8,
+      outputPerToken: 14e-8
+    },
+    "nousresearch/hermes-3-llama-3.1-405b": {
+      inputPerToken: 1e-6,
+      outputPerToken: 1e-6
+    },
+    "nousresearch/hermes-3-llama-3.1-70b": {
+      inputPerToken: 3e-7,
+      outputPerToken: 3e-7
+    },
+    "nousresearch/hermes-4-405b": {
+      inputPerToken: 1e-6,
+      outputPerToken: 3e-6
+    },
+    "nousresearch/hermes-4-70b": {
+      inputPerToken: 13e-8,
+      outputPerToken: 4e-7
+    },
+    "nvidia/llama-3.1-nemotron-70b-instruct": {
+      inputPerToken: 12e-7,
+      outputPerToken: 12e-7
+    },
+    "nvidia/llama-3.3-nemotron-super-49b-v1.5": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "nvidia/nemotron-3-nano-30b-a3b": {
+      inputPerToken: 5e-8,
+      outputPerToken: 2e-7
+    },
+    "nvidia/nemotron-nano-12b-v2-vl": {
+      inputPerToken: 2e-7,
+      outputPerToken: 6e-7
+    },
+    "nvidia/nemotron-nano-9b-v2": {
+      inputPerToken: 4e-8,
+      outputPerToken: 16e-8
+    },
+    "openai/gpt-3.5-turbo": {
+      inputPerToken: 5e-7,
+      outputPerToken: 15e-7
+    },
+    "openai/gpt-3.5-turbo-0613": {
+      inputPerToken: 1e-6,
+      outputPerToken: 2e-6
+    },
+    "openai/gpt-3.5-turbo-16k": {
+      inputPerToken: 3e-6,
+      outputPerToken: 4e-6
+    },
+    "openai/gpt-3.5-turbo-instruct": {
+      inputPerToken: 15e-7,
+      outputPerToken: 2e-6
+    },
+    "openai/gpt-4": {
+      inputPerToken: 3e-5,
+      outputPerToken: 6e-5
+    },
+    "openai/gpt-4-0314": {
+      inputPerToken: 3e-5,
+      outputPerToken: 6e-5
+    },
+    "openai/gpt-4-1106-preview": {
+      inputPerToken: 1e-5,
+      outputPerToken: 3e-5
+    },
+    "openai/gpt-4-turbo": {
+      inputPerToken: 1e-5,
+      outputPerToken: 3e-5
+    },
+    "openai/gpt-4-turbo-preview": {
+      inputPerToken: 1e-5,
+      outputPerToken: 3e-5
+    },
+    "openai/gpt-4.1": {
+      inputPerToken: 2e-6,
+      outputPerToken: 8e-6
+    },
+    "openai/gpt-4.1-mini": {
+      inputPerToken: 4e-7,
+      outputPerToken: 16e-7
+    },
+    "openai/gpt-4.1-nano": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "openai/gpt-4o": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-4o-2024-05-13": {
+      inputPerToken: 5e-6,
+      outputPerToken: 15e-6
+    },
+    "openai/gpt-4o-2024-08-06": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-4o-2024-11-20": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-4o-audio-preview": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-4o-mini": {
+      inputPerToken: 15e-8,
+      outputPerToken: 6e-7
+    },
+    "openai/gpt-4o-mini-2024-07-18": {
+      inputPerToken: 15e-8,
+      outputPerToken: 6e-7
+    },
+    "openai/gpt-4o-mini-search-preview": {
+      inputPerToken: 15e-8,
+      outputPerToken: 6e-7
+    },
+    "openai/gpt-4o-search-preview": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-4o:extended": {
+      inputPerToken: 6e-6,
+      outputPerToken: 18e-6
+    },
+    "openai/gpt-5": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5-chat": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5-codex": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5-image": {
+      inputPerToken: 1e-5,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5-image-mini": {
+      inputPerToken: 25e-7,
+      outputPerToken: 2e-6
+    },
+    "openai/gpt-5-mini": {
+      inputPerToken: 25e-8,
+      outputPerToken: 2e-6
+    },
+    "openai/gpt-5-nano": {
+      inputPerToken: 5e-8,
+      outputPerToken: 4e-7
+    },
+    "openai/gpt-5-pro": {
+      inputPerToken: 15e-6,
+      outputPerToken: 12e-5
+    },
+    "openai/gpt-5.1": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5.1-chat": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5.1-codex": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5.1-codex-max": {
+      inputPerToken: 125e-8,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-5.1-codex-mini": {
+      inputPerToken: 25e-8,
+      outputPerToken: 2e-6
+    },
+    "openai/gpt-5.2": {
+      inputPerToken: 175e-8,
+      outputPerToken: 14e-6
+    },
+    "openai/gpt-5.2-chat": {
+      inputPerToken: 175e-8,
+      outputPerToken: 14e-6
+    },
+    "openai/gpt-5.2-codex": {
+      inputPerToken: 175e-8,
+      outputPerToken: 14e-6
+    },
+    "openai/gpt-5.2-pro": {
+      inputPerToken: 21e-6,
+      outputPerToken: 168e-6
+    },
+    "openai/gpt-5.3-codex": {
+      inputPerToken: 175e-8,
+      outputPerToken: 14e-6
+    },
+    "openai/gpt-audio": {
+      inputPerToken: 25e-7,
+      outputPerToken: 1e-5
+    },
+    "openai/gpt-audio-mini": {
+      inputPerToken: 6e-7,
+      outputPerToken: 24e-7
+    },
+    "openai/gpt-oss-120b": {
+      inputPerToken: 39e-9,
+      outputPerToken: 19e-8
+    },
+    "openai/gpt-oss-120b:exacto": {
+      inputPerToken: 39e-9,
+      outputPerToken: 19e-8
+    },
+    "openai/gpt-oss-20b": {
+      inputPerToken: 3e-8,
+      outputPerToken: 14e-8
+    },
+    "openai/gpt-oss-safeguard-20b": {
+      inputPerToken: 75e-9,
+      outputPerToken: 3e-7
+    },
+    "openai/o1": {
+      inputPerToken: 15e-6,
+      outputPerToken: 6e-5
+    },
+    "openai/o1-pro": {
+      inputPerToken: 15e-5,
+      outputPerToken: 6e-4
+    },
+    "openai/o3": {
+      inputPerToken: 2e-6,
+      outputPerToken: 8e-6
+    },
+    "openai/o3-deep-research": {
+      inputPerToken: 1e-5,
+      outputPerToken: 4e-5
+    },
+    "openai/o3-mini": {
+      inputPerToken: 11e-7,
+      outputPerToken: 44e-7
+    },
+    "openai/o3-mini-high": {
+      inputPerToken: 11e-7,
+      outputPerToken: 44e-7
+    },
+    "openai/o3-pro": {
+      inputPerToken: 2e-5,
+      outputPerToken: 8e-5
+    },
+    "openai/o4-mini": {
+      inputPerToken: 11e-7,
+      outputPerToken: 44e-7
+    },
+    "openai/o4-mini-deep-research": {
+      inputPerToken: 2e-6,
+      outputPerToken: 8e-6
+    },
+    "openai/o4-mini-high": {
+      inputPerToken: 11e-7,
+      outputPerToken: 44e-7
+    },
+    "opengvlab/internvl3-78b": {
+      inputPerToken: 15e-8,
+      outputPerToken: 6e-7
+    },
+    "perplexity/sonar": {
+      inputPerToken: 1e-6,
+      outputPerToken: 1e-6
+    },
+    "perplexity/sonar-deep-research": {
+      inputPerToken: 2e-6,
+      outputPerToken: 8e-6
+    },
+    "perplexity/sonar-pro": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "perplexity/sonar-pro-search": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "perplexity/sonar-reasoning-pro": {
+      inputPerToken: 2e-6,
+      outputPerToken: 8e-6
+    },
+    "prime-intellect/intellect-3": {
+      inputPerToken: 2e-7,
+      outputPerToken: 11e-7
+    },
+    "qwen/qwen-2.5-72b-instruct": {
+      inputPerToken: 12e-8,
+      outputPerToken: 39e-8
+    },
+    "qwen/qwen-2.5-7b-instruct": {
+      inputPerToken: 4e-8,
+      outputPerToken: 1e-7
+    },
+    "qwen/qwen-2.5-coder-32b-instruct": {
+      inputPerToken: 20000000000000002e-23,
+      outputPerToken: 20000000000000002e-23
+    },
+    "qwen/qwen-2.5-vl-7b-instruct": {
+      inputPerToken: 20000000000000002e-23,
+      outputPerToken: 20000000000000002e-23
+    },
+    "qwen/qwen-max": {
+      inputPerToken: 16e-7,
+      outputPerToken: 64e-7
+    },
+    "qwen/qwen-plus": {
+      inputPerToken: 4e-7,
+      outputPerToken: 12e-7
+    },
+    "qwen/qwen-plus-2025-07-28": {
+      inputPerToken: 4e-7,
+      outputPerToken: 12e-7
+    },
+    "qwen/qwen-plus-2025-07-28:thinking": {
+      inputPerToken: 4e-7,
+      outputPerToken: 12e-7
+    },
+    "qwen/qwen-turbo": {
+      inputPerToken: 5e-8,
+      outputPerToken: 2e-7
+    },
+    "qwen/qwen-vl-max": {
+      inputPerToken: 8e-7,
+      outputPerToken: 32e-7
+    },
+    "qwen/qwen-vl-plus": {
+      inputPerToken: 21e-8,
+      outputPerToken: 63e-8
+    },
+    "qwen/qwen2.5-coder-7b-instruct": {
+      inputPerToken: 3e-8,
+      outputPerToken: 9e-8
+    },
+    "qwen/qwen2.5-vl-32b-instruct": {
+      inputPerToken: 2e-7,
+      outputPerToken: 6e-7
+    },
+    "qwen/qwen2.5-vl-72b-instruct": {
+      inputPerToken: 8e-7,
+      outputPerToken: 8e-7
+    },
+    "qwen/qwen3-14b": {
+      inputPerToken: 6e-8,
+      outputPerToken: 24e-8
+    },
+    "qwen/qwen3-235b-a22b": {
+      inputPerToken: 455e-9,
+      outputPerToken: 182e-8
+    },
+    "qwen/qwen3-235b-a22b-2507": {
+      inputPerToken: 71e-9,
+      outputPerToken: 1e-7
+    },
+    "qwen/qwen3-30b-a3b": {
+      inputPerToken: 8e-8,
+      outputPerToken: 28e-8
+    },
+    "qwen/qwen3-30b-a3b-instruct-2507": {
+      inputPerToken: 9e-8,
+      outputPerToken: 3e-7
+    },
+    "qwen/qwen3-30b-a3b-thinking-2507": {
+      inputPerToken: 51e-9,
+      outputPerToken: 34e-8
+    },
+    "qwen/qwen3-32b": {
+      inputPerToken: 8e-8,
+      outputPerToken: 24e-8
+    },
+    "qwen/qwen3-8b": {
+      inputPerToken: 5e-8,
+      outputPerToken: 4e-7
+    },
+    "qwen/qwen3-coder": {
+      inputPerToken: 22e-8,
+      outputPerToken: 1e-6
+    },
+    "qwen/qwen3-coder-30b-a3b-instruct": {
+      inputPerToken: 7e-8,
+      outputPerToken: 27e-8
+    },
+    "qwen/qwen3-coder-flash": {
+      inputPerToken: 3e-7,
+      outputPerToken: 15e-7
+    },
+    "qwen/qwen3-coder-next": {
+      inputPerToken: 12e-8,
+      outputPerToken: 75e-8
+    },
+    "qwen/qwen3-coder-plus": {
+      inputPerToken: 1e-6,
+      outputPerToken: 5e-6
+    },
+    "qwen/qwen3-coder:exacto": {
+      inputPerToken: 22e-8,
+      outputPerToken: 18e-7
+    },
+    "qwen/qwen3-max": {
+      inputPerToken: 12e-7,
+      outputPerToken: 6e-6
+    },
+    "qwen/qwen3-max-thinking": {
+      inputPerToken: 12e-7,
+      outputPerToken: 6e-6
+    },
+    "qwen/qwen3-next-80b-a3b-instruct": {
+      inputPerToken: 9e-8,
+      outputPerToken: 11e-7
+    },
+    "qwen/qwen3-next-80b-a3b-thinking": {
+      inputPerToken: 15e-8,
+      outputPerToken: 12e-7
+    },
+    "qwen/qwen3-vl-235b-a22b-instruct": {
+      inputPerToken: 2e-7,
+      outputPerToken: 88e-8
+    },
+    "qwen/qwen3-vl-30b-a3b-instruct": {
+      inputPerToken: 13e-8,
+      outputPerToken: 52e-8
+    },
+    "qwen/qwen3-vl-32b-instruct": {
+      inputPerToken: 104e-9,
+      outputPerToken: 416e-9
+    },
+    "qwen/qwen3-vl-8b-instruct": {
+      inputPerToken: 8e-8,
+      outputPerToken: 5e-7
+    },
+    "qwen/qwen3-vl-8b-thinking": {
+      inputPerToken: 117e-9,
+      outputPerToken: 1365e-9
+    },
+    "qwen/qwen3.5-122b-a10b": {
+      inputPerToken: 4e-7,
+      outputPerToken: 32e-7
+    },
+    "qwen/qwen3.5-27b": {
+      inputPerToken: 3e-7,
+      outputPerToken: 24e-7
+    },
+    "qwen/qwen3.5-35b-a3b": {
+      inputPerToken: 25e-8,
+      outputPerToken: 2e-6
+    },
+    "qwen/qwen3.5-397b-a17b": {
+      inputPerToken: 55e-8,
+      outputPerToken: 35e-7
+    },
+    "qwen/qwen3.5-flash-02-23": {
+      inputPerToken: 1e-7,
+      outputPerToken: 4e-7
+    },
+    "qwen/qwen3.5-plus-02-15": {
+      inputPerToken: 4e-7,
+      outputPerToken: 24e-7
+    },
+    "qwen/qwq-32b": {
+      inputPerToken: 15e-8,
+      outputPerToken: 4e-7
+    },
+    "raifle/sorcererlm-8x22b": {
+      inputPerToken: 45e-7,
+      outputPerToken: 45e-7
+    },
+    "relace/relace-apply-3": {
+      inputPerToken: 85e-8,
+      outputPerToken: 125e-8
+    },
+    "relace/relace-search": {
+      inputPerToken: 1e-6,
+      outputPerToken: 3e-6
+    },
+    "sao10k/l3-euryale-70b": {
+      inputPerToken: 148e-8,
+      outputPerToken: 148e-8
+    },
+    "sao10k/l3-lunaris-8b": {
+      inputPerToken: 4e-8,
+      outputPerToken: 5e-8
+    },
+    "sao10k/l3.1-70b-hanami-x1": {
+      inputPerToken: 3e-6,
+      outputPerToken: 3e-6
+    },
+    "sao10k/l3.1-euryale-70b": {
+      inputPerToken: 65e-8,
+      outputPerToken: 75e-8
+    },
+    "sao10k/l3.3-euryale-70b": {
+      inputPerToken: 65e-8,
+      outputPerToken: 75e-8
+    },
+    "stepfun/step-3.5-flash": {
+      inputPerToken: 1e-7,
+      outputPerToken: 3e-7
+    },
+    "switchpoint/router": {
+      inputPerToken: 85e-8,
+      outputPerToken: 34e-7
+    },
+    "tencent/hunyuan-a13b-instruct": {
+      inputPerToken: 14e-8,
+      outputPerToken: 57e-8
+    },
+    "thedrummer/cydonia-24b-v4.1": {
+      inputPerToken: 3e-7,
+      outputPerToken: 5e-7
+    },
+    "thedrummer/rocinante-12b": {
+      inputPerToken: 17e-8,
+      outputPerToken: 43e-8
+    },
+    "thedrummer/skyfall-36b-v2": {
+      inputPerToken: 55e-8,
+      outputPerToken: 8e-7
+    },
+    "thedrummer/unslopnemo-12b": {
+      inputPerToken: 4e-7,
+      outputPerToken: 4e-7
+    },
+    "tngtech/deepseek-r1t2-chimera": {
+      inputPerToken: 25e-8,
+      outputPerToken: 85e-8
+    },
+    "undi95/remm-slerp-l2-13b": {
+      inputPerToken: 45e-8,
+      outputPerToken: 65e-8
+    },
+    "writer/palmyra-x5": {
+      inputPerToken: 6e-7,
+      outputPerToken: 6e-6
+    },
+    "xai/grok-3": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "xai/grok-3-beta": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "xai/grok-3-mini": {
+      inputPerToken: 3e-7,
+      outputPerToken: 5e-7
+    },
+    "xai/grok-3-mini-beta": {
+      inputPerToken: 3e-7,
+      outputPerToken: 5e-7
+    },
+    "xai/grok-4": {
+      inputPerToken: 3e-6,
+      outputPerToken: 15e-6
+    },
+    "xai/grok-4-fast": {
+      inputPerToken: 2e-7,
+      outputPerToken: 5e-7
+    },
+    "xai/grok-4.1-fast": {
+      inputPerToken: 2e-7,
+      outputPerToken: 5e-7
+    },
+    "xai/grok-code-fast-1": {
+      inputPerToken: 2e-7,
+      outputPerToken: 15e-7
+    },
+    "xiaomi/mimo-v2-flash": {
+      inputPerToken: 9e-8,
+      outputPerToken: 29e-8
+    },
+    "z-ai/glm-4-32b": {
+      inputPerToken: 1e-7,
+      outputPerToken: 1e-7
+    },
+    "z-ai/glm-4.5": {
+      inputPerToken: 55e-8,
+      outputPerToken: 2e-6
+    },
+    "z-ai/glm-4.5-air": {
+      inputPerToken: 13e-8,
+      outputPerToken: 85e-8
+    },
+    "z-ai/glm-4.5v": {
+      inputPerToken: 6e-7,
+      outputPerToken: 18e-7
+    },
+    "z-ai/glm-4.6": {
+      inputPerToken: 35e-8,
+      outputPerToken: 171e-8
+    },
+    "z-ai/glm-4.6:exacto": {
+      inputPerToken: 44e-8,
+      outputPerToken: 176e-8
+    },
+    "z-ai/glm-4.6v": {
+      inputPerToken: 3e-7,
+      outputPerToken: 9e-7
+    },
+    "z-ai/glm-4.7": {
+      inputPerToken: 3e-7,
+      outputPerToken: 14e-7
+    },
+    "z-ai/glm-4.7-flash": {
+      inputPerToken: 6e-8,
+      outputPerToken: 4e-7
+    },
+    "z-ai/glm-5": {
+      inputPerToken: 95e-8,
+      outputPerToken: 255e-8
+    }
+  }
+};
+
+// src/pricing/lookup.ts
+var models = catalog_default.models;
+var modelNameIndex = /* @__PURE__ */ new Map();
+for (const key of Object.keys(models)) {
+  const name = key.split("/").slice(1).join("/");
+  if (name && !modelNameIndex.has(name)) {
+    modelNameIndex.set(name, key);
+  }
+}
+function lookupPricing(providerId) {
+  if (models[providerId]) return models[providerId];
+  const model = providerId.split("/").slice(1).join("/");
+  if (!model) return void 0;
+  const asOpenai = `openai/${model}`;
+  if (models[asOpenai]) return models[asOpenai];
+  const crossKey = modelNameIndex.get(model);
+  if (crossKey) return models[crossKey];
+  return void 0;
+}
+function estimateCost(pricing, promptTokens, completionTokens) {
+  return pricing.inputPerToken * promptTokens + pricing.outputPerToken * completionTokens;
+}
+
+// src/scorers/cost.ts
+var costScorer = ({ result }, providerId) => {
+  const promptTokens = result.usage?.promptTokens ?? 0;
+  const completionTokens = result.usage?.completionTokens ?? 0;
+  const totalTokens = promptTokens + completionTokens;
+  const pricing = lookupPricing(providerId);
+  if (!pricing) {
+    return {
+      name: "cost",
+      value: -1,
+      details: {
+        estimatedUsd: null,
+        promptTokens,
+        completionTokens,
+        totalTokens,
+        note: "No pricing data available for this model"
+      }
+    };
+  }
+  const usd = estimateCost(pricing, promptTokens, completionTokens);
+  return {
+    name: "cost",
+    value: usd,
+    details: {
+      estimatedUsd: usd,
+      promptTokens,
+      completionTokens,
+      totalTokens
+    }
+  };
+};
+
+// src/utils/deep-equal.ts
+function deepEqual(expected, actual) {
+  if (expected === actual) return true;
+  if (typeof expected === "string" && typeof actual === "string") {
+    return expected.trim().toLowerCase() === actual.trim().toLowerCase();
+  }
+  if (typeof expected !== typeof actual) return false;
+  if (expected === null || actual === null) return expected === actual;
+  if (Array.isArray(expected) && Array.isArray(actual)) {
+    if (expected.length !== actual.length) return false;
+    return expected.every((val, i7) => deepEqual(val, actual[i7]));
+  }
+  if (typeof expected === "object" && typeof actual === "object") {
+    const objExpected = expected;
+    const objActual = actual;
+    const keysExpected = Object.keys(objExpected);
+    return keysExpected.every((key) => key in objActual && deepEqual(objExpected[key], objActual[key]));
+  }
+  return expected === actual;
+}
+
+// src/scorers/correctness.ts
+var correctnessScorer = ({ task, result }) => {
+  if (task.expected === void 0) {
+    return { name: "correctness", value: 0.5, details: { reason: "no expected value" } };
+  }
+  const actual = normalizeOutput(task.expected, result.output);
+  const match = deepEqual(task.expected, actual);
+  return {
+    name: "correctness",
+    value: match ? 1 : 0,
+    details: { expected: task.expected, actual: result.output }
+  };
+};
+function normalizeOutput(expected, actual) {
+  if (Array.isArray(expected) && !Array.isArray(actual) && typeof actual === "object" && actual !== null) {
+    const entries = Object.entries(actual);
+    const arrayEntries = entries.filter(([, v4]) => Array.isArray(v4));
+    if (arrayEntries.length === 1) {
+      return arrayEntries[0][1];
+    }
+  }
+  return actual;
+}
+
+// src/scorers/schema-correctness.ts
+var schemaCorrectnessScorer = ({ task, result }) => {
+  if (!task.schema) {
+    return { name: "schema-correctness", value: -1, details: { reason: "no schema defined" } };
+  }
+  let data = result.output;
+  if (typeof data === "string") {
+    try {
+      data = JSON.parse(data);
+    } catch {
+      return {
+        name: "schema-correctness",
+        value: 0,
+        details: { reason: "output is not valid JSON" }
+      };
+    }
+  }
+  let parsed = task.schema.safeParse(data);
+  if (!parsed.success && !Array.isArray(data) && typeof data === "object" && data !== null) {
+    const arrayEntries = Object.entries(data).filter(([, v4]) => Array.isArray(v4));
+    if (arrayEntries.length === 1) {
+      const unwrapped = task.schema.safeParse(arrayEntries[0][1]);
+      if (unwrapped.success) parsed = unwrapped;
+    }
+  }
+  return {
+    name: "schema-correctness",
+    value: parsed.success ? 1 : 0,
+    details: parsed.success ? { valid: true } : { valid: false, errors: parsed.error.issues.map((i7) => i7.message) }
+  };
+};
+
+// src/scorers/fuzzy-similarity.ts
+var fuzzySimilarityScorer = ({ task, result }) => {
+  if (task.expected === void 0) {
+    return { name: "fuzzy-similarity", value: -1, details: { reason: "no expected value" } };
+  }
+  const a7 = stringify(task.expected);
+  const b3 = stringify(result.output);
+  const setA = tokenize(a7);
+  const setB = tokenize(b3);
+  const similarity = jaccardSimilarity(setA, setB);
+  return {
+    name: "fuzzy-similarity",
+    value: Math.round(similarity * 100) / 100,
+    details: { method: "jaccard", expectedTokens: setA.size, actualTokens: setB.size }
+  };
+};
+function stringify(value) {
+  if (typeof value === "string") return value.toLowerCase();
+  return JSON.stringify(value).toLowerCase();
+}
+function tokenize(text) {
+  return new Set(text.match(/\w+/g) ?? []);
+}
+function jaccardSimilarity(a7, b3) {
+  if (a7.size === 0 && b3.size === 0) return 1;
+  let intersection = 0;
+  for (const token of a7) {
+    if (b3.has(token)) intersection++;
+  }
+  const union = a7.size + b3.size - intersection;
+  return union === 0 ? 1 : intersection / union;
+}
+
+// src/scorers/llm-judge.ts
+import OpenAI2, { AzureOpenAI as AzureOpenAI2 } from "openai";
+
+// src/providers/openai.ts
+import OpenAI, { AzureOpenAI } from "openai";
+import { zodToJsonSchema as zodToJsonSchema2 } from "zod-to-json-schema";
+
+// src/providers/shared.ts
+import { zodToJsonSchema } from "zod-to-json-schema";
+
+// src/providers/openai.ts
+var REQUEST_TIMEOUT_MS = 6e4;
+
+// src/scorers/llm-judge.ts
+var JUDGE_PROMPT = `You are a strict scoring judge. Evaluate the actual output against the expected output on three criteria. Score each from 0.0 to 1.0 using the full range (not just 0, 0.5, 1).
+
+Criteria:
+1. Accuracy \u2014 are the facts, entities, and claims correct? Penalize hallucinations or wrong details.
+2. Completeness \u2014 does it capture all key information from the expected output? Penalize missing points.
+3. Conciseness \u2014 is it free of unnecessary filler, repetition, or tangential content? Penalize verbosity.
+
+Respond with ONLY this exact format \u2014 three lines, no other text:
+accuracy: <number>
+completeness: <number>
+conciseness: <number>
+
+Task: {task}
+Expected: {expected}
+Actual: {actual}`;
+function resolveJudgeClient(configModel, timeoutMs = REQUEST_TIMEOUT_MS) {
+  const model = configModel ?? process.env.DUELIST_JUDGE_MODEL ?? "gpt-5-mini";
+  if (model.startsWith("gemini") && process.env.GOOGLE_API_KEY) {
+    return {
+      client: new OpenAI2({
+        apiKey: process.env.GOOGLE_API_KEY,
+        baseURL: "https://generativelanguage.googleapis.com/v1beta/openai/",
+        timeout: timeoutMs
+      }),
+      model
+    };
+  }
+  if (!process.env.OPENAI_API_KEY && process.env.AZURE_OPENAI_API_KEY) {
+    return {
+      client: new AzureOpenAI2({
+        apiKey: process.env.AZURE_OPENAI_API_KEY,
+        endpoint: process.env.AZURE_OPENAI_ENDPOINT,
+        apiVersion: process.env.AZURE_OPENAI_API_VERSION ?? "2024-12-01-preview",
+        deployment: model,
+        timeout: timeoutMs
+      }),
+      model
+    };
+  }
+  const apiKey = process.env.OPENAI_API_KEY;
+  if (!apiKey) return void 0;
+  return { client: new OpenAI2({ apiKey, timeout: timeoutMs }), model };
+}
+function isTemperatureError(err) {
+  const msg = err instanceof Error ? err.message : String(err);
+  const lower = msg.toLowerCase();
+  return lower.includes("temperature") && (lower.includes("not supported") || lower.includes("is not allowed") || lower.includes("unsupported") || lower.includes("invalid"));
+}
+function createLlmJudgeScorer(judgeModel, timeoutMs = REQUEST_TIMEOUT_MS) {
+  let cached = void 0;
+  let useTemperature = true;
+  return async ({ task, result }) => {
+    if (task.expected === void 0) {
+      return { name: "llm-judge-correctness", value: -1, details: { reason: "no expected value" } };
+    }
+    if (cached === void 0) {
+      cached = resolveJudgeClient(judgeModel, timeoutMs) ?? null;
+    }
+    if (!cached) {
+      return {
+        name: "llm-judge-correctness",
+        value: -1,
+        details: { reason: "no API key available for judge model" }
+      };
+    }
+    const { client, model } = cached;
+    const prompt = JUDGE_PROMPT.replace("{task}", task.prompt).replace("{expected}", JSON.stringify(task.expected)).replace("{actual}", JSON.stringify(result.output));
+    const messages = [{ role: "user", content: prompt }];
+    try {
+      const response = await callJudge(client, model, messages, useTemperature);
+      return parseJudgeResponse(response, model);
+    } catch (err) {
+      if (useTemperature && isTemperatureError(err)) {
+        useTemperature = false;
+        try {
+          const response = await callJudge(client, model, messages, false);
+          return parseJudgeResponse(response, model);
+        } catch (retryErr) {
+          return {
+            name: "llm-judge-correctness",
+            value: -1,
+            details: { reason: `judge call failed: ${retryErr instanceof Error ? retryErr.message : String(retryErr)}` }
+          };
+        }
+      }
+      return {
+        name: "llm-judge-correctness",
+        value: -1,
+        details: { reason: `judge call failed: ${err instanceof Error ? err.message : String(err)}` }
+      };
+    }
+  };
+}
+async function callJudge(client, model, messages, withTemperature) {
+  return client.chat.completions.create({
+    model,
+    messages,
+    max_completion_tokens: 2048,
+    ...withTemperature ? { temperature: 0 } : {}
+  });
+}
+function parseJudgeResponse(response, model) {
+  const content = response.choices[0]?.message?.content?.trim() ?? "";
+  const parsed = {};
+  for (const line of content.split("\n")) {
+    const match = line.match(/^(accuracy|completeness|conciseness)\s*:\s*([\d.]+)/i);
+    if (match) parsed[match[1].toLowerCase()] = parseFloat(match[2]);
+  }
+  const accuracy = parsed.accuracy;
+  const completeness = parsed.completeness;
+  const conciseness = parsed.conciseness;
+  if (accuracy == null || completeness == null || conciseness == null || [accuracy, completeness, conciseness].some((s5) => isNaN(s5) || s5 < 0 || s5 > 1)) {
+    return {
+      name: "llm-judge-correctness",
+      value: -1,
+      details: { reason: `judge returned unparseable scores: "${content}"`, model }
+    };
+  }
+  const composite = Math.round((accuracy + completeness + conciseness) / 3 * 100) / 100;
+  return {
+    name: "llm-judge-correctness",
+    value: composite,
+    details: { model, accuracy, completeness, conciseness }
+  };
+}
+
+// src/scorers/tool-usage.ts
+var toolUsageScorer = ({ task, result }) => {
+  if (!task.tools?.length) {
+    return { name: "tool-usage", value: -1, details: { reason: "no tools configured on task" } };
+  }
+  const calls = result.toolCalls ?? [];
+  const expectedIsObject = task.expected !== void 0 && typeof task.expected === "object" && task.expected !== null && !Array.isArray(task.expected);
+  if (expectedIsObject) {
+    const matchingCall = calls.find((c3) => {
+      const toolDef = task.tools.find((t3) => t3.name === c3.name);
+      if (!toolDef) return false;
+      return deepEqual(task.expected, c3.arguments);
+    });
+    if (matchingCall) {
+      return {
+        name: "tool-usage",
+        value: 1,
+        details: { matchedTool: matchingCall.name, arguments: matchingCall.arguments, toolCalls: calls }
+      };
+    }
+    const expectedKeys = Object.keys(task.expected);
+    const partialMatch = calls.find((c3) => {
+      if (typeof c3.arguments !== "object" || c3.arguments === null) return false;
+      const argKeys = Object.keys(c3.arguments);
+      return expectedKeys.some((k3) => argKeys.includes(k3));
+    });
+    if (partialMatch) {
+      return {
+        name: "tool-usage",
+        value: 0.5,
+        details: {
+          reason: "correct tool but wrong arguments",
+          expected: task.expected,
+          actual: partialMatch.arguments,
+          toolCalls: calls
+        }
+      };
+    }
+    return {
+      name: "tool-usage",
+      value: 0,
+      details: { reason: "no matching tool call", expected: task.expected, toolCalls: calls }
+    };
+  }
+  const expectedToolName = task.tools[0].name;
+  const usedTool = calls.some((c3) => c3.name === expectedToolName);
+  return {
+    name: "tool-usage",
+    value: usedTool ? 1 : 0,
+    details: { expectedToolName, usedTool, toolCalls: calls }
+  };
+};
+
+// src/scorers/index.ts
+var staticScorers = {
+  latency: latencyScorer,
+  cost: costScorer,
+  correctness: correctnessScorer,
+  "schema-correctness": schemaCorrectnessScorer,
+  "fuzzy-similarity": fuzzySimilarityScorer,
+  "tool-usage": toolUsageScorer
+};
+function resolveScorers(names, judgeModel, timeoutMs) {
+  return names.map((name) => {
+    if (name === "llm-judge-correctness") {
+      return createLlmJudgeScorer(judgeModel, timeoutMs);
+    }
+    const scorer = staticScorers[name];
+    if (!scorer) {
+      throw new Error(`Unknown scorer: "${name}"`);
+    }
+    return scorer;
+  });
+}
+
+// src/runner.ts
+var DEFAULT_TIMEOUT_MS = 6e4;
+function withTimeout(run, ms) {
+  return new Promise((resolve2, reject) => {
+    const controller = new AbortController();
+    const timer = setTimeout(() => {
+      controller.abort();
+      reject(new Error(`Request timed out after ${ms}ms`));
+    }, ms);
+    run(controller.signal).then(
+      (v4) => {
+        clearTimeout(timer);
+        resolve2(v4);
+      },
+      (e5) => {
+        clearTimeout(timer);
+        reject(e5);
+      }
+    );
+  });
+}
+async function runBenchmarks(options) {
+  const { providers, tasks, scorers, runs, onResult } = options;
+  const timeout = options.timeout ?? DEFAULT_TIMEOUT_MS;
+  const results = [];
+  for (const task of tasks) {
+    for (let run = 1; run <= runs; run++) {
+      const runResults = await Promise.all(
+        providers.map(async (provider) => {
+          let result;
+          try {
+            const taskResult = await withTimeout((signal) => provider.run({
+              prompt: task.prompt,
+              schema: task.schema,
+              tools: task.tools,
+              signal,
+              timeout
+            }), timeout);
+            const scores = await Promise.all(
+              scorers.map((scorer) => scorer({ task, result: taskResult }, provider.id))
+            );
+            result = {
+              providerId: provider.id,
+              taskName: task.name,
+              run,
+              scores,
+              raw: {
+                output: taskResult.output,
+                latencyMs: taskResult.latencyMs,
+                usage: taskResult.usage,
+                toolCalls: taskResult.toolCalls
+              }
+            };
+          } catch (err) {
+            const message = err instanceof Error ? err.message : String(err);
+            result = {
+              providerId: provider.id,
+              taskName: task.name,
+              run,
+              scores: [],
+              error: message,
+              raw: { output: "", latencyMs: 0 }
+            };
+          }
+          onResult?.(result);
+          return result;
+        })
+      );
+      results.push(...runResults);
+    }
+  }
+  return results;
+}
+
 // src/utils/format.ts
 var MAX_FRACTION_DIGITS = 100;
 function formatCost(usd) {
@@ -6368,37 +8110,76 @@ function computeColumnStats(providerData, scorerNames) {
   }
   return stats;
 }
+var QUALITY_SCORERS = /* @__PURE__ */ new Set([
+  "correctness",
+  "schema-correctness",
+  "fuzzy-similarity",
+  "llm-judge-correctness",
+  "tool-usage"
+]);
+function passesQualityGate(providerId, columnStats) {
+  const qualityColumns = [...columnStats.keys()].filter((k3) => QUALITY_SCORERS.has(k3));
+  if (qualityColumns.length === 0) return true;
+  return qualityColumns.some((col) => {
+    const val = columnStats.get(col)?.values.get(providerId);
+    return val !== void 0 && val > 0;
+  });
+}
 function computeMedals(columnStats, providerIds) {
   const medals = /* @__PURE__ */ new Map();
   if (providerIds.length < 2) {
     for (const id of providerIds) medals.set(id, "none");
     return medals;
   }
-  const wins = /* @__PURE__ */ new Map();
-  for (const id of providerIds) wins.set(id, 0);
-  for (const [, colStats] of columnStats) {
+  const eligible = new Set(providerIds.filter((id) => passesQualityGate(id, columnStats)));
+  const qualityWins = /* @__PURE__ */ new Map();
+  const efficiencyWins = /* @__PURE__ */ new Map();
+  for (const id of providerIds) {
+    qualityWins.set(id, 0);
+    efficiencyWins.set(id, 0);
+  }
+  for (const [colName, colStats] of columnStats) {
     if (colStats.best === void 0) continue;
     const bestProviders = [...colStats.values.entries()].filter(([, v4]) => v4 !== void 0 && v4 === colStats.best);
     if (bestProviders.length === 1) {
-      wins.set(bestProviders[0][0], (wins.get(bestProviders[0][0]) ?? 0) + 1);
+      const winnerId = bestProviders[0][0];
+      if (QUALITY_SCORERS.has(colName)) {
+        qualityWins.set(winnerId, (qualityWins.get(winnerId) ?? 0) + 1);
+      } else {
+        efficiencyWins.set(winnerId, (efficiencyWins.get(winnerId) ?? 0) + 1);
+      }
     }
   }
-  const totalWins = [...wins.values()].reduce((a7, b3) => a7 + b3, 0);
+  const totalWins = [...qualityWins.values()].reduce((a7, b3) => a7 + b3, 0) + [...efficiencyWins.values()].reduce((a7, b3) => a7 + b3, 0);
   if (totalWins === 0) {
     for (const id of providerIds) medals.set(id, "none");
     return medals;
   }
-  const sorted = [...wins.entries()].sort(
-    (a7, b3) => b3[1] - a7[1] || a7[0].localeCompare(b3[0])
-  );
+  const eligibleSorted = providerIds.filter((id) => eligible.has(id)).sort((a7, b3) => {
+    const qDiff = (qualityWins.get(b3) ?? 0) - (qualityWins.get(a7) ?? 0);
+    if (qDiff !== 0) return qDiff;
+    const eDiff = (efficiencyWins.get(b3) ?? 0) - (efficiencyWins.get(a7) ?? 0);
+    if (eDiff !== 0) return eDiff;
+    return a7.localeCompare(b3);
+  });
   const medalList = ["gold", "silver", "bronze"];
   let rank = 0;
-  for (let i7 = 0; i7 < sorted.length; i7++) {
-    if (i7 > 0 && sorted[i7][1] < sorted[i7 - 1][1]) {
-      rank = i7;
+  for (let i7 = 0; i7 < eligibleSorted.length; i7++) {
+    if (i7 > 0) {
+      const prevQ = qualityWins.get(eligibleSorted[i7 - 1]) ?? 0;
+      const currQ = qualityWins.get(eligibleSorted[i7]) ?? 0;
+      if (currQ < prevQ) {
+        rank = i7;
+      } else if (currQ === prevQ) {
+        const prevE = efficiencyWins.get(eligibleSorted[i7 - 1]) ?? 0;
+        const currE = efficiencyWins.get(eligibleSorted[i7]) ?? 0;
+        if (currE < prevE) rank = i7;
+      }
     }
-    const hasWins = sorted[i7][1] > 0;
-    medals.set(sorted[i7][0], hasWins && rank < medalList.length ? medalList[rank] : "none");
+    medals.set(eligibleSorted[i7], rank < medalList.length ? medalList[rank] : "none");
+  }
+  for (const id of providerIds) {
+    if (!eligible.has(id)) medals.set(id, "none");
   }
   return medals;
 }
@@ -6795,24 +8576,10 @@ function printSummary(results, providers, byProvider) {
       console.log(`  ${medal} Cheapest:      ${bold(byCost.id)} ${dim(providerLabel(byCost.id))}  ${brightGreen}${boldCode}${costStr}${reset}`);
     }
   }
-  if (!single) {
-    const wins = /* @__PURE__ */ new Map();
-    for (const id of providers) wins.set(id, 0);
-    if (byCorrectness) wins.set(byCorrectness.id, (wins.get(byCorrectness.id) ?? 0) + 1);
-    if (byLatency && byLatency.avg !== Infinity) wins.set(byLatency.id, (wins.get(byLatency.id) ?? 0) + 1);
-    if (byCost?.avg !== void 0) wins.set(byCost.id, (wins.get(byCost.id) ?? 0) + 1);
-    const maxWins = Math.max(...wins.values());
-    if (maxWins > 0) {
-      const topProviders = [...wins.entries()].filter(([, w4]) => w4 === maxWins);
-      console.log("");
-      if (topProviders.length === 1) {
-        const [winnerId, winCount] = topProviders[0];
-        console.log(`  \u{1F3C6} Overall:      ${brightGreen}${boldCode}${winnerId}${reset} ${dim(providerLabel(winnerId))}  ${dim(`(${winCount}/3 categories)`)}`);
-      } else {
-        const names = topProviders.map(([id]) => bold(id)).join(dim(", "));
-        console.log(`  \u{1F3C6} Overall:      ${names}  ${dim(`(tied at ${maxWins}/3)`)}`);
-      }
-    }
+  if (!single && byCorrectness && byCorrectness.avg > 0) {
+    console.log("");
+    const pct = `${Math.round(byCorrectness.avg * 100)}%`;
+    console.log(`  \u{1F3C6} Overall:      ${brightGreen}${boldCode}${byCorrectness.id}${reset} ${dim(providerLabel(byCorrectness.id))}  ${dim(`(${pct} avg correctness)`)}`);
   }
   console.log("");
 }
@@ -6841,6 +8608,32 @@ function buildSummary(results) {
   };
 }
 
+// src/arena.ts
+function defineArena(config) {
+  if (config.providers.length === 0) {
+    throw new Error("At least one provider is required");
+  }
+  const scorerNames = config.scorers ?? ["latency", "cost", "correctness"];
+  const scorerFns = resolveScorers(scorerNames, config.judgeModel, config.timeout);
+  const runs = config.runs ?? 1;
+  return {
+    config,
+    async run(options) {
+      if (config.tasks.length === 0) {
+        throw new Error("At least one task is required");
+      }
+      return runBenchmarks({
+        providers: config.providers,
+        tasks: config.tasks,
+        scorers: scorerFns,
+        runs,
+        timeout: config.timeout,
+        onResult: options?.onResult
+      });
+    }
+  };
+}
+
 // src/reporter/markdown.ts
 var COMMENT_MARKER = "<!-- duelist-ci-report -->";
 function markdownReporter(report, _current) {
@@ -6968,17 +8761,8 @@ function htmlReporter(results) {
     return { id, avg };
   }).filter((p5) => p5.avg !== void 0).sort((a7, b3) => a7.avg - b3.avg)[0];
   let overallWinner;
-  if (multi) {
-    const wins = /* @__PURE__ */ new Map();
-    for (const id of providers) wins.set(id, 0);
-    if (byCorrectness) wins.set(byCorrectness.id, (wins.get(byCorrectness.id) ?? 0) + 1);
-    if (byLatency && byLatency.avg !== Infinity) wins.set(byLatency.id, (wins.get(byLatency.id) ?? 0) + 1);
-    if (byCost?.avg !== void 0) wins.set(byCost.id, (wins.get(byCost.id) ?? 0) + 1);
-    const maxWins = Math.max(...wins.values());
-    if (maxWins > 0) {
-      const tops = [...wins.entries()].filter(([, w4]) => w4 === maxWins);
-      if (tops.length === 1) overallWinner = tops[0][0];
-    }
+  if (multi && byCorrectness && byCorrectness.avg > 0) {
+    overallWinner = byCorrectness.id;
   }
   const errorResults = results.filter((r3) => r3.error);
   const deduped = dedupeErrors(errorResults);
@@ -7499,7 +9283,7 @@ function renderErrors(errors) {
     </div>`;
   }).join("\n");
   return `<section class="errors-section">
-  <h2 class="errors-title" onclick="this.nextElementSibling.style.display=this.nextElementSibling.style.display==='none'?'block':'block'">Errors</h2>
+  <h2 class="errors-title" onclick="this.nextElementSibling.style.display=this.nextElementSibling.style.display==='none'?'block':'none'">Errors</h2>
   <div class="errors-list">
     ${items}
   </div>
@@ -7843,6 +9627,403 @@ async function upsertPrComment(ctx, body, marker) {
   }
 }
 
+// src/packs/structured-output.ts
+import { z } from "zod";
+var structuredOutputPack = {
+  name: "structured-output",
+  label: "Structured Output",
+  description: "Zod schema stress test \u2014 flat objects, nesting, arrays, enums, empty arrays, and adversarial input",
+  tasks: [
+    {
+      name: "so:flat-entity",
+      prompt: "Extract the person's details from this text: 'Maria Garcia, age 34, works as a software architect in Barcelona, Spain. Her employee ID is EMP-2847.' Return as JSON.",
+      expected: {
+        name: "Maria Garcia",
+        age: 34,
+        role: "software architect",
+        city: "Barcelona",
+        country: "Spain",
+        employeeId: "EMP-2847"
+      },
+      schema: z.object({
+        name: z.string(),
+        age: z.number(),
+        role: z.string(),
+        city: z.string(),
+        country: z.string(),
+        employeeId: z.string()
+      })
+    },
+    {
+      name: "so:nested-object",
+      prompt: "Parse this shipping label into structured JSON: 'Ship to: Acme Corp, Attn: John Lee, 4th Floor, 742 Evergreen Terrace, Springfield, IL 62704, USA. Order #ORD-9912, 3 items, 2.4kg, express shipping.' Use shippingMethod values: standard, express, or overnight. Return as JSON.",
+      expected: {
+        recipient: { company: "Acme Corp", contact: "John Lee", floor: "4th Floor" },
+        address: { street: "742 Evergreen Terrace", city: "Springfield", state: "IL", zip: "62704", country: "USA" },
+        order: { id: "ORD-9912", itemCount: 3, weightKg: 2.4, shippingMethod: "express" }
+      },
+      schema: z.object({
+        recipient: z.object({ company: z.string(), contact: z.string(), floor: z.string() }),
+        address: z.object({
+          street: z.string(),
+          city: z.string(),
+          state: z.string(),
+          zip: z.string(),
+          country: z.string()
+        }),
+        order: z.object({
+          id: z.string(),
+          itemCount: z.number(),
+          weightKg: z.number(),
+          shippingMethod: z.enum(["standard", "express", "overnight"])
+        })
+      })
+    },
+    {
+      name: "so:array-of-objects",
+      prompt: "Extract all mentioned products with their prices and categories from this text: 'Our summer sale includes the UltraWidget Pro ($49.99, Electronics), ComfortMax Chair ($199.00, Furniture), and AquaPure Filter ($24.50, Home & Kitchen). The SmartLamp Mini is also available at $34.99 in the Electronics category.' Return as a JSON array.",
+      expected: [
+        { name: "UltraWidget Pro", price: 49.99, category: "Electronics" },
+        { name: "ComfortMax Chair", price: 199, category: "Furniture" },
+        { name: "AquaPure Filter", price: 24.5, category: "Home & Kitchen" },
+        { name: "SmartLamp Mini", price: 34.99, category: "Electronics" }
+      ],
+      schema: z.array(z.object({ name: z.string(), price: z.number(), category: z.string() }))
+    },
+    {
+      name: "so:empty-arrays",
+      prompt: "Extract all error codes and their severity levels from this log message: 'System health check completed at 14:32 UTC. All services operational. No warnings or errors detected. Uptime: 99.97%.' Classify status as one of: healthy, degraded, or down. Return as JSON.",
+      expected: { errors: [], warnings: [], status: "healthy", uptimePercent: 99.97 },
+      schema: z.object({
+        errors: z.array(z.object({ code: z.string(), severity: z.string() })),
+        warnings: z.array(z.string()),
+        status: z.enum(["healthy", "degraded", "down"]),
+        uptimePercent: z.number()
+      })
+    },
+    {
+      name: "so:enum-classification",
+      prompt: "Classify each of these support tickets by priority (low/medium/high/critical) and category (billing/technical/account/general). Use just the letter (A, B, C, D) as the id.\nTicket A: 'My account was charged twice for the same subscription.'\nTicket B: 'The API returns 500 errors intermittently.'\nTicket C: 'How do I update my display name?'\nTicket D: 'Production database is completely unresponsive, all services down.'\nReturn as a JSON array.",
+      expected: [
+        { id: "A", priority: "high", category: "billing" },
+        { id: "B", priority: "high", category: "technical" },
+        { id: "C", priority: "low", category: "account" },
+        { id: "D", priority: "critical", category: "technical" }
+      ],
+      schema: z.array(
+        z.object({
+          id: z.string(),
+          priority: z.enum(["low", "medium", "high", "critical"]),
+          category: z.enum(["billing", "technical", "account", "general"])
+        })
+      )
+    },
+    {
+      name: "so:adversarial-input",
+      prompt: `Extract the actual product review data from this messy input. Ignore any JSON-like noise in the text.
+
+User said: 'I bought the {product: "fake"} headphones for $59.99 and they're great! Rating: 5/5. The "noise-cancelling" feature works well even in {"noisy": true} environments. Would recommend to friend=true. Purchased on 01/15/2026.'
+Return as JSON. Use ISO 8601 date format (YYYY-MM-DD).`,
+      expected: {
+        product: "headphones",
+        price: 59.99,
+        rating: 5,
+        maxRating: 5,
+        features: ["noise-cancelling"],
+        recommended: true,
+        purchaseDate: "2026-01-15"
+      },
+      schema: z.object({
+        product: z.string(),
+        price: z.number(),
+        rating: z.number(),
+        maxRating: z.number(),
+        features: z.array(z.string()),
+        recommended: z.boolean(),
+        purchaseDate: z.string()
+      })
+    }
+  ],
+  scorers: ["correctness", "schema-correctness", "latency", "cost"]
+};
+
+// src/packs/tool-calling.ts
+import { z as z2 } from "zod";
+var toolCallingPack = {
+  name: "tool-calling",
+  label: "Tool Calling",
+  description: "Function invocation accuracy \u2014 single calls, complex params, tool selection, parallel calls, and relevance detection",
+  tasks: [
+    {
+      name: "tc:simple-single-tool",
+      prompt: "What's the current weather in Tokyo?",
+      tools: [{
+        name: "getWeather",
+        description: "Get current weather for a city",
+        parameters: z2.object({
+          city: z2.string(),
+          units: z2.enum(["celsius", "fahrenheit"]).optional()
+        }),
+        handler: async ({ city, units }) => ({
+          city,
+          tempC: 8,
+          condition: "cloudy",
+          units: units ?? "celsius"
+        })
+      }],
+      expected: { city: "Tokyo" }
+    },
+    {
+      name: "tc:complex-params",
+      prompt: "Search for Italian restaurants within 2 miles of downtown Portland that are open now and have at least a 4-star rating.",
+      tools: [{
+        name: "searchRestaurants",
+        description: "Search for restaurants matching criteria",
+        parameters: z2.object({
+          cuisine: z2.string(),
+          location: z2.string(),
+          radiusMiles: z2.number(),
+          minRating: z2.number(),
+          openNow: z2.boolean()
+        }),
+        handler: async (_args) => ({
+          results: [{ name: "Trattoria Roma", rating: 4.5, distance: 1.2 }]
+        })
+      }],
+      expected: {
+        cuisine: "Italian",
+        location: "downtown Portland",
+        radiusMiles: 2,
+        minRating: 4,
+        openNow: true
+      }
+    },
+    {
+      name: "tc:select-from-many",
+      prompt: "Convert 150 USD to Euros.",
+      tools: [
+        {
+          name: "getWeather",
+          description: "Get current weather for a city",
+          parameters: z2.object({ city: z2.string() }),
+          handler: async () => ({ tempC: 20 })
+        },
+        {
+          name: "convertCurrency",
+          description: "Convert an amount between currencies",
+          parameters: z2.object({
+            amount: z2.number(),
+            from: z2.string(),
+            to: z2.string()
+          }),
+          handler: async ({ amount, from, to }) => ({
+            amount,
+            from,
+            to,
+            result: 138.75,
+            rate: 0.925
+          })
+        },
+        {
+          name: "translateText",
+          description: "Translate text between languages",
+          parameters: z2.object({ text: z2.string(), targetLang: z2.string() }),
+          handler: async () => ({ translated: "" })
+        },
+        {
+          name: "calculateTip",
+          description: "Calculate tip amount for a bill",
+          parameters: z2.object({ billAmount: z2.number(), tipPercent: z2.number() }),
+          handler: async () => ({ tip: 0 })
+        }
+      ],
+      expected: { amount: 150, from: "USD", to: "EUR" }
+    },
+    {
+      name: "tc:parallel-calls",
+      prompt: "I'm planning a trip. What's the weather like in both Paris and London right now?",
+      tools: [{
+        name: "getWeather",
+        description: "Get current weather for a city",
+        parameters: z2.object({ city: z2.string() }),
+        handler: async ({ city }) => {
+          const data = {
+            Paris: { tempC: 12, condition: "partly cloudy" },
+            London: { tempC: 9, condition: "rainy" }
+          };
+          return data[city] ?? { tempC: 15, condition: "unknown" };
+        }
+      }],
+      expected: "weather data for Paris and London"
+    }
+  ],
+  scorers: ["tool-usage", "latency", "cost"]
+};
+
+// src/packs/reasoning.ts
+import { z as z3 } from "zod";
+var reasoningPack = {
+  name: "reasoning",
+  label: "Reasoning",
+  description: "Logic, math, and multi-step thinking \u2014 arithmetic, deduction, data interpretation, critical path, and business rules",
+  tasks: [
+    {
+      name: "rs:saas-mrr-calc",
+      prompt: `A SaaS company charges $49/month for the basic plan and $149/month for pro.
+In Q1 they had 200 basic subscribers and 85 pro subscribers.
+In Q2, 15% of basic users upgraded to pro and they gained 40 new basic subscribers.
+No one churned. What is the Q2 monthly recurring revenue (MRR)?
+Return as JSON with your reasoning and the final MRR number.`,
+      expected: { mrr: 27425 },
+      schema: z3.object({
+        reasoning: z3.string().optional(),
+        mrr: z3.number()
+      })
+    },
+    {
+      name: "rs:logical-deduction",
+      prompt: `Five developers \u2014 Alice, Bob, Carol, Dave, and Eve \u2014 each use a different
+primary language: Rust, TypeScript, Python, Go, and Java. Given:
+1. Alice does not use Python, Java, or Go.
+2. Bob uses TypeScript.
+3. Carol uses neither Rust nor Go.
+4. Dave does not use Java.
+5. Eve uses neither Rust, Go, nor Java.
+What language does each developer use? Return as JSON.`,
+      expected: {
+        Alice: "Rust",
+        Bob: "TypeScript",
+        Carol: "Java",
+        Dave: "Go",
+        Eve: "Python"
+      },
+      schema: z3.object({
+        Alice: z3.string(),
+        Bob: z3.string(),
+        Carol: z3.string(),
+        Dave: z3.string(),
+        Eve: z3.string()
+      })
+    },
+    {
+      name: "rs:data-interpretation",
+      prompt: `Given this quarterly revenue data:
+| Quarter | Revenue | Growth |
+|---------|---------|--------|
+| Q1 2025 | $2.1M   | -      |
+| Q2 2025 | $2.4M   | 14.3%  |
+| Q3 2025 | $2.2M   | -8.3%  |
+| Q4 2025 | $2.8M   | 27.3%  |
+
+Which quarter had the highest absolute revenue increase compared to the previous
+quarter? What was the full-year total revenue in millions? Return as JSON.`,
+      expected: {
+        highestGrowthQuarter: "Q4 2025",
+        absoluteIncrease: 0.6,
+        fullYearRevenue: 9.5
+      },
+      schema: z3.object({
+        highestGrowthQuarter: z3.string(),
+        absoluteIncrease: z3.number(),
+        fullYearRevenue: z3.number()
+      })
+    },
+    {
+      name: "rs:critical-path",
+      prompt: `A deployment pipeline has these stages with dependencies:
+- Build (3 min, no dependency)
+- Unit tests (5 min, depends on Build)
+- Integration tests (8 min, depends on Build)
+- Security scan (4 min, depends on Build)
+- Staging deploy (2 min, depends on Unit tests AND Integration tests AND Security scan)
+- Smoke tests (3 min, depends on Staging deploy)
+
+Assuming stages run in parallel where possible, what is the total pipeline
+duration in minutes? Which stages are on the critical path? Return as JSON.`,
+      expected: {
+        totalMinutes: 16,
+        criticalPath: ["Build", "Integration tests", "Staging deploy", "Smoke tests"]
+      },
+      schema: z3.object({
+        totalMinutes: z3.number(),
+        criticalPath: z3.array(z3.string())
+      })
+    },
+    {
+      name: "rs:pricing-rules",
+      prompt: `Apply these pricing rules to each customer and return the final price:
+Rules:
+- Base price: $100
+- Enterprise customers (>100 seats): 30% discount
+- Annual billing: additional 15% off the discounted price
+- Non-profit organizations: flat $50 regardless of other rules
+
+Customers:
+A: 50 seats, monthly billing, for-profit
+B: 200 seats, annual billing, for-profit
+C: 75 seats, annual billing, non-profit
+D: 150 seats, monthly billing, for-profit
+
+Return as a JSON array with customer id and finalPrice.`,
+      expected: [
+        { id: "A", finalPrice: 100 },
+        { id: "B", finalPrice: 59.5 },
+        { id: "C", finalPrice: 50 },
+        { id: "D", finalPrice: 70 }
+      ],
+      schema: z3.array(z3.object({
+        id: z3.string(),
+        finalPrice: z3.number()
+      }))
+    }
+  ],
+  scorers: ["correctness", "latency", "cost"]
+};
+
+// src/packs/index.ts
+var registry = /* @__PURE__ */ new Map();
+function register(pack) {
+  registry.set(pack.name, pack);
+}
+register(structuredOutputPack);
+register(toolCallingPack);
+register(reasoningPack);
+function loadPack(name) {
+  const pack = registry.get(name);
+  if (!pack) {
+    const available = [...registry.keys()].join(", ");
+    throw new Error(`Unknown pack "${name}". Available packs: ${available}`);
+  }
+  return pack;
+}
+function listPacks() {
+  return [...registry.values()].map((p5) => ({
+    name: p5.name,
+    label: p5.label,
+    description: p5.description,
+    taskCount: p5.tasks.length
+  }));
+}
+
+// src/packs/loader.ts
+function buildPackConfig(config) {
+  const packs = config.packs.map((name) => loadPack(name));
+  const tasks = packs.flatMap((p5) => p5.tasks);
+  const scorerSet = /* @__PURE__ */ new Set();
+  for (const pack of packs) {
+    for (const scorer of pack.scorers) {
+      scorerSet.add(scorer);
+    }
+  }
+  return {
+    providers: config.providers,
+    tasks,
+    scorers: [...scorerSet],
+    runs: config.runs ?? 1,
+    timeout: config.timeout
+  };
+}
+
 // src/cli.ts
 var __dirname2 = dirname2(fileURLToPath(import.meta.url));
 var program = new Command();
@@ -7867,12 +10048,16 @@ program.command("init").description("Scaffold an arena.config.ts in the current
   console.log("  1. export OPENAI_API_KEY=sk-...");
   console.log("  2. npx duelist run");
 });
-program.command("run").description("Run benchmarks defined in your arena config").option("-c, --config <path>", "Path to config file", "arena.config.ts").option("--reporter <type>", "Output format: console, json, or html", "console").option("--output <path>", "Output file path (used with html reporter)", "duelist-report.html").option("-q, --quiet", "Suppress per-result progress (show only final report)").action(async (opts) => {
+program.command("run").description("Run benchmarks defined in your arena config").option("-c, --config <path>", "Path to config file", "arena.config.ts").option("--pack <names>", 'Run built-in task pack(s), comma-separated. Use "list" to see available packs.').option("--reporter <type>", "Output format: console, json, or html", "console").option("--output <path>", "Output file path (used with html reporter)", "duelist-report.html").option("-q, --quiet", "Suppress per-result progress (show only final report)").action(async (opts) => {
+  if (opts.pack === "list") {
+    printPackList();
+    return;
+  }
   if (!["console", "json", "html"].includes(opts.reporter)) {
     console.error(`Unknown reporter "${opts.reporter}". Use "console", "json", or "html".`);
     process.exit(1);
   }
-  const typedArena = await loadArenaConfig(opts.config);
+  const typedArena = opts.pack ? await loadArenaWithPacks(opts.pack, opts.config) : await loadArenaConfig(opts.config);
   try {
     const showProgress = opts.reporter !== "json" && !opts.quiet;
     const onResult = showProgress ? logResult : void 0;
@@ -7907,7 +10092,11 @@ function collectThreshold(value, previous) {
   previous.set(scorer, Number(delta));
   return previous;
 }
-program.command("ci").description("Run benchmarks, compare against baseline, and enforce quality gates").option("-c, --config <path>", "Path to config file", "arena.config.ts").option("--baseline <path>", "Baseline JSON file", ".duelist/baseline.json").option("--budget <dollars>", "Max total cost in USD", parseFloat).option("--threshold <scorer=delta>", "Regression threshold (repeatable)", collectThreshold, /* @__PURE__ */ new Map()).option("--update-baseline", "Save results as new baseline after passing").option("--comment", "Post results as GitHub PR comment").option("-q, --quiet", "Suppress per-result progress").action(async (opts) => {
+program.command("ci").description("Run benchmarks, compare against baseline, and enforce quality gates").option("-c, --config <path>", "Path to config file", "arena.config.ts").option("--pack <names>", 'Run built-in task pack(s), comma-separated. Use "list" to see available packs.').option("--baseline <path>", "Baseline JSON file", ".duelist/baseline.json").option("--budget <dollars>", "Max total cost in USD", parseFloat).option("--threshold <scorer=delta>", "Regression threshold (repeatable)", collectThreshold, /* @__PURE__ */ new Map()).option("--update-baseline", "Save results as new baseline after passing").option("--comment", "Post results as GitHub PR comment").option("-q, --quiet", "Suppress per-result progress").action(async (opts) => {
+  if (opts.pack === "list") {
+    printPackList();
+    return;
+  }
   const ciOpts = {
     configPath: opts.config,
     baselinePath: resolve(opts.baseline),
@@ -7917,7 +10106,7 @@ program.command("ci").description("Run benchmarks, compare against baseline, and
     comment: opts.comment ?? false,
     quiet: opts.quiet ?? false
   };
-  const typedArena = await loadArenaConfig(ciOpts.configPath);
+  const typedArena = opts.pack ? await loadArenaWithPacks(opts.pack, ciOpts.configPath) : await loadArenaConfig(ciOpts.configPath);
   console.log("Running benchmarks...");
   const onResult = ciOpts.quiet ? void 0 : logResult;
   let results;
@@ -7974,6 +10163,39 @@ program.command("ci").description("Run benchmarks, compare against baseline, and
   process.exit(report.failed ? 1 : 0);
 });
 program.parse();
+function printPackList() {
+  const packs = listPacks();
+  if (packs.length === 0) {
+    console.log("No packs available.");
+    return;
+  }
+  const nameWidth = Math.max(...packs.map((p5) => p5.name.length)) + 2;
+  console.log("Available task packs:\n");
+  for (const p5 of packs) {
+    const tasks = `${p5.taskCount} tasks`;
+    console.log(`  ${p5.name.padEnd(nameWidth)} ${tasks.padEnd(9)} ${p5.description}`);
+  }
+  console.log(`
+Run:     npx duelist run --pack <name>`);
+  console.log(`Combine: npx duelist run --pack structured-output,tool-calling`);
+}
+async function loadArenaWithPacks(packNames, configOpt) {
+  const configPath = resolve(configOpt);
+  if (!existsSync(configPath)) {
+    console.error("No arena.config.ts found. Create one with `npx duelist init` to configure");
+    console.error("your providers, then re-run with --pack.");
+    process.exit(1);
+  }
+  const userArena = await loadArenaConfig(configOpt);
+  const packs = packNames.split(",").map((s5) => s5.trim());
+  const packConfig = buildPackConfig({
+    packs,
+    providers: userArena.config.providers,
+    runs: userArena.config.runs,
+    timeout: userArena.config.timeout
+  });
+  return defineArena(packConfig);
+}
 async function loadArenaConfig(configOpt) {
   const configPath = resolve(configOpt);
   if (!existsSync(configPath)) {