@kenkaiiii/gg-core 5.80.1 → 5.82.0

package/dist/model-registry.cjs

@@ -49,6 +49,7 @@ __export(model_registry_exports, {
   usesOpenAICodexTransport: () => usesOpenAICodexTransport
 });
 module.exports = __toCommonJS(model_registry_exports);
+var import_gg_ai2 = require("@kenkaiiii/gg-ai");
 
 // src/oauth/kimi.ts
 var import_node_child_process = require("child_process");
@@ -170,9 +171,8 @@ var MODELS = [
     // context, 128K output, image input. Same always-on adaptive thinking with
     // the full effort ladder (low→max, xhigh included; the API-side default
     // effort is `medium`) and the same Fable 5.1 constraints gg-ai already
-    // handles (no forced tool use, thinking can't be disabled). Also eligible
-    // for the gated `speed: "fast"` research preview (waitlist-only, premium
-    // $8/$40 pricing — not wired up). Opus 5 is retired here — a session that
+    // handles (no forced tool use, thinking can't be disabled). Fast mode is
+    // the separate `claude-opus-5-5-fast` entry below. Opus 5 is retired here — a session that
     // still has it saved falls back to the provider default on next start.
     id: "claude-opus-5-5",
     name: "Claude Opus 5.5",
@@ -189,6 +189,25 @@ var MODELS = [
     defaultThinkingLevel: "medium",
     maxThinkingLevel: "max"
   },
+  {
+    // Opus 5.5 in fast mode (research preview): the same model at up to 2.5x
+    // output tokens/sec, $8/$40 MTok. gg-ai's `stream()` sends it as
+    // `claude-opus-5-5` + `speed: "fast"` + the fast-mode beta. On a Claude
+    // subscription it bills usage credits only — never included plan usage —
+    // and Anthropic 429s "Usage credits are required for fast mode" until they
+    // are turned on. Its prompt cache is separate from standard-speed Opus.
+    id: "claude-opus-5-5-fast",
+    name: "Claude Opus 5.5 Fast",
+    provider: "anthropic",
+    contextWindow: 1e6,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "high",
+    defaultThinkingLevel: "medium",
+    maxThinkingLevel: "max"
+  },
   {
     // Released 2026-09-28 — replaces Sonnet 5 at $2/$10 MTok, with the same
     // 1M context / 128K output and adaptive thinking, now including xhigh.
@@ -256,6 +275,24 @@ var MODELS = [
     costTier: "high",
     maxThinkingLevel: "ultra"
   },
+  {
+    // Codex Fast mode for Astra (catalog `service_tiers: priority`, "2x speed,
+    // increased usage"). gg-ai's `stream()` sends it as `gpt-6-astra` +
+    // `service_tier: "priority"`, exactly what the Codex CLI sends. On a
+    // ChatGPT plan it draws included limits at 2.5x the standard rate.
+    id: "gpt-6-astra-fast",
+    name: "GPT-6 Astra Fast",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "low",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "high",
+    maxThinkingLevel: "ultra"
+  },
   // GPT-6 Sol + Luna — released 2026-09-22 below Astra, replacing the whole
   // GPT-5.6 family (Sol/Terra/Luna; there is no GPT-6 Terra — OpenAI's Codex
   // catalog upgrades 5.6 Terra to 6 Sol). GPT-6.1 Sol (2026-09-29) then replaced
@@ -284,6 +321,22 @@ var MODELS = [
     costTier: "medium",
     maxThinkingLevel: "ultra"
   },
+  {
+    // Codex Fast mode for 6.1 Sol ("2x speed, increased usage") — see
+    // `gpt-6-astra-fast` for the wire mapping and billing.
+    id: "gpt-6.1-sol-fast",
+    name: "GPT-6.1 Sol Fast",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "low",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "medium",
+    maxThinkingLevel: "ultra"
+  },
   {
     // Luna — "Fast and affordable model for easier tasks." (Codex priority 4,
     // default medium, needs a Codex client >= 0.155.0). $0.10/$0.50 MTok.
@@ -301,6 +354,23 @@ var MODELS = [
     costTier: "low",
     maxThinkingLevel: "max"
   },
+  {
+    // Codex Fast mode for Luna ("1.5x speed") — see `gpt-6-astra-fast` for the
+    // wire mapping and billing. Listed after Luna so compaction summaries
+    // (first `low` OpenAI model) stay on standard-speed Luna.
+    id: "gpt-6-luna-fast",
+    name: "GPT-6 Luna Fast",
+    provider: "openai",
+    contextWindow: 105e4,
+    codexContextWindow: 272e3,
+    maxOutputTokens: 128e3,
+    supportsThinking: true,
+    defaultThinkingLevel: "medium",
+    supportsImages: true,
+    supportsVideo: false,
+    costTier: "low",
+    maxThinkingLevel: "max"
+  },
   // ── Sakana (Fugu) ──────────────────────────────────────
   // Sakana Fugu is a multi-agent system surfaced as a standard LLM via the
   // OpenAI-compatible Sakana API (https://api.sakana.ai/v1). All three take
@@ -874,7 +944,9 @@ var SUMMARY_COST_TIER = {
 };
 function getSummaryModel(provider, currentModelId) {
   const tier = SUMMARY_COST_TIER[provider];
-  const tagged = tier ? getModelsForProvider(provider).find((m) => m.costTier === tier) : void 0;
+  const tagged = tier ? getModelsForProvider(provider).find(
+    (m) => m.costTier === tier && !(0, import_gg_ai2.resolveFastModeModel)(provider, m.id).fast
+  ) : void 0;
   return tagged ?? getModel(currentModelId) ?? getDefaultModel(provider);
 }
 // Annotate the CommonJS export names for ESM import in node: