@hiper2d/ai-agents 0.8.0 → 0.10.0

package/dist/index.js

@@ -1065,11 +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,
@@ -1084,8 +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,
@@ -1125,10 +1129,15 @@ var SupportedAiModels = {
     reasoningEffort: "minimal",
     tags: ["fast", "cheap"]
   },
-  // Always-on reasoning (xAI default effort "high", cannot be disabled) — no non-thinking sibling
+  // Always-on reasoning — no non-thinking sibling. Bumped 4.6 → 4.7 on 2026-09-22 (id verified
+  // against GET /v1/models). Same price, same 200K extended-context boundary; the context
+  // window is 500K. NOTE 4.7 newly accepts `reasoning_effort` (low | medium | high | xhigh,
+  // default high) where 4.6 had no knob at all. GrokAgent still sends none, so behaviour is
+  // unchanged — but Grok is the app's most expensive model per turn (94% of its output is
+  // reasoning), so a lower pin is the obvious lever to measure before anything else.
   [LLM_CONSTANTS.GROK]: {
-    displayName: "Grok 4.6",
-    modelApiName: "grok-4.6",
+    displayName: "Grok 4.7",
+    modelApiName: "grok-4.7",
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
     temperature: 0.7
@@ -1397,17 +1406,20 @@ 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]: {
@@ -1420,12 +1432,12 @@ 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
@@ -1537,8 +1549,9 @@ var MODEL_PRICING = {
     cacheHitPrice: 0.015
   },
   // Grok models. Cached price is per-model on xAI (not a uniform ratio):
-  // grok-4.6 is $0.50/M cached vs $2.00/M input, and all rates double for prompts
-  // >= 200K tokens, per docs.x.ai/developers/models (verified 2026-08-12).
+  // grok-4.7 is $0.50/M cached vs $2.00/M input, and all rates double for prompts
+  // >= 200K tokens, per docs.x.ai/docs/models (re-verified 2026-09-22 for 4.7).
+  // 4.7 carries over 4.6's rates unchanged — the bump is capability-only, not a reprice.
   [SupportedAiModels[LLM_CONSTANTS.GROK].modelApiName]: {
     inputPrice: 2,
     outputPrice: 6,