@hiper2d/ai-agents 0.7.0 → 0.9.0

package/dist/index.js

@@ -998,9 +998,13 @@ var SupportedAiModels = {
     tags: ["expensive"]
   },
   // Claude models — thinking-only entries (non-thinking variants retired 2026-08-05)
+  // Opus moved 5 → 5.5 on 2026-09-22 (`claude-opus-5-5`, verified against GET /v1/models).
+  // Cheaper than Opus 5 ($4/$20 vs $5/$25) and the same 1M context / 128K output. Thinking is
+  // always on, as on Opus 5. Its API default effort is `medium`, but we keep the catalog-wide
+  // `high` pin so bot answers don't get shallower than every other model in the lobby.
   [LLM_CONSTANTS.CLAUDE_OPUS]: {
-    displayName: "Claude 5 Opus",
-    modelApiName: "claude-opus-5",
+    displayName: "Claude 5.5 Opus",
+    modelApiName: "claude-opus-5-5",
     apiKeyName: API_KEY_CONSTANTS.ANTHROPIC,
     hasThinking: true,
     reasoningEffort: "high",
@@ -1061,11 +1065,15 @@ var SupportedAiModels = {
     temperature: 1,
     tags: ["expensive"]
   },
-  // GPT-5.6 family (promoted July 2026 when the limited preview opened up):
-  // sol is the flagship, terra the mainline, luna the cheap tier.
+  // Sol and Luna moved to GPT-6 on 2026-09-22 (both ids verified against GET /v1/models).
+  // GPT-6 ships Astra, Sol and Luna only — there is NO gpt-6-terra, so the Terra slot stays on
+  // gpt-5.6-terra, which OpenAI still serves. Both moves are big price cuts: Sol halved
+  // ($4/$20 → $2/$10) and Luna halved ($0.20/$1.20 → $0.10/$0.50). Note GPT-6 Sol now
+  // undercuts GPT-5.6 Terra ($2/$12) on output at the same input rate, which makes the Terra
+  // slot largely redundant — retiring it is a product call, not a catalog one.
   [LLM_CONSTANTS.GPT_SOL]: {
-    displayName: "GPT-5.6 Sol",
-    modelApiName: "gpt-5.6-sol",
+    displayName: "GPT-6 Sol",
+    modelApiName: "gpt-6-sol",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
     hasThinking: true,
     temperature: 1,
@@ -1080,8 +1088,8 @@ var SupportedAiModels = {
     tags: ["fast", "expensive"]
   },
   [LLM_CONSTANTS.GPT_MINI]: {
-    displayName: "GPT-5.6 Luna",
-    modelApiName: "gpt-5.6-luna",
+    displayName: "GPT-6 Luna",
+    modelApiName: "gpt-6-luna",
     apiKeyName: API_KEY_CONSTANTS.OPENAI,
     hasThinking: true,
     temperature: 1,
@@ -1393,17 +1401,20 @@ var MODEL_PRICING = {
     extendedContextCacheHitPrice: 2,
     extendedContextThresholdTokens: 272e3
   },
-  // OpenAI GPT-5.6 models
-  // Sol repriced 2026-08-30 (developers.openai.com/api/docs/pricing): $4/$20 short context,
-  // $8/$30 past the long-context threshold — the same 272k boundary its siblings use.
-  // Cache writes ($5/$10) are not modelled; OpenAI caching is automatic and we only see hits.
+  // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22),
+  // plus GPT-5.6 Terra, which has no GPT-6 successor and keeps its old rates.
+  // Sol: $2/$10 short, $4/$15 long, cache hits $0.20/$0.40 — half what GPT-5.6 Sol cost.
+  // Luna: $0.10/$0.50 short, $0.20/$0.75 long, cache hits $0.01/$0.02 — also halved.
+  // As with Astra, OpenAI's table doesn't restate the short/long boundary, so we keep the
+  // 272k threshold the 5.6 siblings use. Cache WRITES (Sol $2.50/$5, Luna $0.125/$0.25) are
+  // not modelled: OpenAI caching is automatic and the API only reports hits.
   [SupportedAiModels[LLM_CONSTANTS.GPT_SOL].modelApiName]: {
-    inputPrice: 4,
-    outputPrice: 20,
-    cacheHitPrice: 0.4,
-    extendedContextInputPrice: 8,
-    extendedContextOutputPrice: 30,
-    extendedContextCacheHitPrice: 0.8,
+    inputPrice: 2,
+    outputPrice: 10,
+    cacheHitPrice: 0.2,
+    extendedContextInputPrice: 4,
+    extendedContextOutputPrice: 15,
+    extendedContextCacheHitPrice: 0.4,
     extendedContextThresholdTokens: 272e3
   },
   [SupportedAiModels[LLM_CONSTANTS.GPT].modelApiName]: {
@@ -1416,12 +1427,12 @@ var MODEL_PRICING = {
     extendedContextThresholdTokens: 272e3
   },
   [SupportedAiModels[LLM_CONSTANTS.GPT_MINI].modelApiName]: {
-    inputPrice: 0.2,
-    outputPrice: 1.2,
-    cacheHitPrice: 0.02,
-    extendedContextInputPrice: 0.4,
-    extendedContextOutputPrice: 1.8,
-    extendedContextCacheHitPrice: 0.04,
+    inputPrice: 0.1,
+    outputPrice: 0.5,
+    cacheHitPrice: 0.01,
+    extendedContextInputPrice: 0.2,
+    extendedContextOutputPrice: 0.75,
+    extendedContextCacheHitPrice: 0.02,
     extendedContextThresholdTokens: 272e3
   },
   // DeepSeek V4 models
@@ -1473,9 +1484,13 @@ var MODEL_PRICING = {
     cacheHitPrice: 0.25
   },
   [SupportedAiModels[LLM_CONSTANTS.CLAUDE_OPUS].modelApiName]: {
-    inputPrice: 5,
-    outputPrice: 25,
-    cacheHitPrice: 0.5
+    // Opus 5.5 (2026-09-22): $4/$20, down from Opus 5's $5/$25. Cache reads are 5% of input
+    // here, not the usual 10% — $0.20, not $0.40. Cache WRITES ($5 at 5m, $8 at 1h) are not
+    // modelled: calculateCost bills every uncached prompt token at inputPrice, so a write is
+    // under-billed by the 1.25x/2x premium. Pre-existing for every Anthropic model.
+    inputPrice: 4,
+    outputPrice: 20,
+    cacheHitPrice: 0.2
   },
   [SupportedAiModels[LLM_CONSTANTS.CLAUDE_SONNET].modelApiName]: {
     inputPrice: 2,
@@ -2633,6 +2648,7 @@ ${instruction}`);
 
 // src/agents/anthropic-agent.ts
 var import_sdk = require("@anthropic-ai/sdk");
+var rejectsDisabledThinking = (model) => model.includes("fable") || model.includes("opus-5-5");
 var ClaudeAgent = class extends AbstractAgent {
   client;
   /**
@@ -2849,7 +2865,9 @@ ${schemaDescription}`;
           params.temperature = 1;
         }
       } else if (usesAdaptiveThinking) {
-        params.thinking = { type: "disabled" };
+        if (!rejectsDisabledThinking(this.model)) {
+          params.thinking = { type: "disabled" };
+        }
       } else {
         params.temperature = this.temperature;
       }
@@ -2943,7 +2961,9 @@ ${schemaDescription}`;
           params.temperature = 1;
         }
       } else if (usesAdaptiveThinking) {
-        params.thinking = { type: "disabled" };
+        if (!rejectsDisabledThinking(this.model)) {
+          params.thinking = { type: "disabled" };
+        }
       } else {
         params.temperature = this.temperature;
       }