@molecule/api-resource-ai-models 1.0.0

package/dist/models.js

@@ -0,0 +1,1233 @@
+/**
+ * Available AI models.
+ *
+ * This is the single source of truth for which models Synthase can use.
+ * Server-side code (chat handler, compaction) consumes the full definitions
+ * directly; clients receive the `PublicModel` projection from `GET /ai/models`.
+ *
+ * @module
+ */
+/**
+ * All available AI models, grouped by provider, ordered from most to least capable.
+ *
+ * To add or remove a model, edit this array. Both the server-side validation
+ * and the public discovery endpoint will update automatically.
+ *
+ * Effort is each model's OWN native value — there is no abstract scale (see
+ * {@link ModelDefinition.supportedEffortLevels}):
+ * - A model driven by a provider-native effort/level param lists its provider
+ *   values verbatim in `supportedEffortLevels` (ascending), with
+ *   `defaultEffortLevel` = the provider's default/recommended level for agentic
+ *   coding. NO `effortBudgetTokens`.
+ * - A model with a controllable token budget but no native level names (e.g.
+ *   Claude Haiku 4.5's `budget_tokens`, Qwen's `thinking_budget`) lists
+ *   scaled-budget labels (`['4K', '8K', '16K', '32K']`) with `effortBudgetTokens`
+ *   mapping each label to the token budget it sends.
+ * - A model whose reasoning is fixed (always-on or on/off only, no depth
+ *   control) carries `thinkingConfigurable: false` and OMITS both fields —
+ *   there is nothing to tune.
+ *
+ * Sources (verified 2026-07-28; OpenAI re-verified 2026-07-31 after the
+ * 2026-07-30 GPT-5.6 repricing — cross-check prices against models.dev with
+ * `npm run check:model-freshness` from the workspace root):
+ * - Anthropic: https://platform.claude.com/docs/en/about-claude/models/overview
+ *   + /docs/en/build-with-claude/effort (fable-5 / opus-5 / sonnet-5 current;
+ *   opus-4-8 superseded by opus-5 at identical pricing but still served — it is
+ *   the recommended refusal-fallback model; effort ladder on all three current
+ *   models is low|medium|high|xhigh|max; budget_tokens 400s on 4.7+)
+ * - OpenAI: https://developers.openai.com/api/docs/pricing (GPT-5.6 family GA
+ *   2026-07-09; REPRICED 2026-07-30: -luna cut 80% to $0.20/$1.20, -terra cut
+ *   20% to $2/$12, -sol unchanged $5/$30; cache read 0.1× input; gpt-5.5/
+ *   gpt-5.4 still listed as current; long-context 2× variants exist upstream —
+ *   not modeled, same as the Gemini/Grok tiers; Sol "Fast mode" 2.5× speed at
+ *   2× price announced 2026-07-30 — not yet modeled)
+ * - Google: https://ai.google.dev/gemini-api/docs/pricing (gemini-3.6-flash GA
+ *   2026-07-21 $1.50/$7.50 supersedes 3.5-flash as the agentic flagship;
+ *   gemini-3.1-pro-preview still the pro tier — "3.5 Pro" has NOT shipped as
+ *   of 2026-07-28 despite the coming-soon badge; do not add until it has an id)
+ * - xAI: https://docs.x.ai/developers/models + /developers/grok-4-5
+ *   (grok-4.5 flagship 2026-07-08: $2/$6, 500K ctx, ≥200K prompts bill 2× —
+ *   not modeled; reasoning_effort low|medium|high default high, image input;
+ *   grok-4.3 still served at $1.25/$2.50 with the bigger 1M window;
+ *   grok-code-fast-1 no longer listed — retires 2026-08-15)
+ * - DeepSeek: https://api-docs.deepseek.com/quick_start/pricing (unchanged V4
+ *   Pro/Flash pricing; legacy deepseek-chat/-reasoner ids fully retired
+ *   2026-07-24 — never in this catalog; the announced peak-hour 2× surcharge is
+ *   still NOT active as of 2026-07-28, see the entries)
+ * - Moonshot: https://platform.kimi.ai/docs/models (kimi-k3 flagship 2026-07-16
+ *   — 2.8T MoE, 1M ctx, $3/$15 — NOT added: thinking is forced-on with
+ *   reasoning_content that must be replayed through tool loops, the same
+ *   constraint that keeps kimi-k2.7-code out; add BOTH once the moonshot bond
+ *   supports preserved thinking + reasoning_effort low|high|max. kimi-k2.6
+ *   remains the newest model the bond can run correctly.)
+ * - MiniMax: https://platform.minimax.io/docs/guides/pricing-paygo (unchanged;
+ *   minimax-m3 $0.30/$1.20 is a "permanent 50% off" list rate)
+ * - Alibaba: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+ *   (unchanged; qwen3.8-max-preview, 2026-07-19, is Token-Plan-subscriber-only
+ *   — not on the pay-as-you-go API, so it cannot be added yet; qwen3.7-max
+ *   currently runs a 50%-off promo — still billed here at list, $2.50/$7.50)
+ * - Zhipu: https://docs.z.ai/guides/overview/pricing (unchanged; glm-5.2 is
+ *   the newest — "GLM-5.3/5.5" rumors have no released ids as of 2026-07-28)
+ *
+ * Knowledge-cutoff dates on non-Anthropic entries are best-effort estimates
+ * where the provider doesn't publish one; the provider sources above verify
+ * id / pricing / context window.
+ */
+export const MODELS = [
+    // ---------------------------------------------------------------------------
+    // Anthropic
+    // Verified: https://platform.claude.com/docs/en/about-claude/models/overview
+    //           https://platform.claude.com/docs/en/build-with-claude/effort
+    // Effort is output_config.effort with adaptive thinking on 4.6+ models —
+    // budget_tokens is REJECTED (400) on fable-5 / opus-4-8 / sonnet-5 and
+    // deprecated on the 4.6 family. Only Haiku 4.5 still uses budget_tokens.
+    // ---------------------------------------------------------------------------
+    {
+        id: 'claude-fable-5',
+        provider: 'anthropic',
+        label: 'Claude Fable 5',
+        description: 'Most capable Anthropic — frontier reasoning & long-horizon agents',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+        defaultEffortLevel: 'high',
+        // Thinking is ALWAYS ON (adaptive); effort is the only depth control.
+        // Full five-level ladder (medium was missing here — Fable supports low
+        // through max). Default/recommended is high; xhigh for the most
+        // capability-sensitive work; low still performs well on routine tasks.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        codeExecutionToolType: 'code_execution_20250825',
+        webFetchToolType: 'web_fetch_20260209',
+        inputPricePerMTok: 10,
+        outputPricePerMTok: 50,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 1,
+        cacheWritePricePerMTok: 12.5,
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'claude-opus-5',
+        provider: 'anthropic',
+        label: 'Claude Opus 5',
+        description: 'Anthropic Opus flagship — step-change agentic coding at 4.8 pricing',
+        // Fast mode (research preview, Claude API only): same model at up to 2.5×
+        // output speed, $10/$50 per MTok (2× standard; platform.claude.com fast-mode
+        // docs, verified 2026-07-31). Cache rates follow Anthropic's standard
+        // ratios (read 0.1× input, write 1.25× input) applied to the fast input
+        // rate. Separate upstream rate limits from standard Opus.
+        fastPricing: {
+            inputPricePerMTok: 10,
+            outputPricePerMTok: 50,
+            cacheReadPricePerMTok: 1,
+            cacheWritePricePerMTok: 12.5,
+        },
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+        defaultEffortLevel: 'high',
+        // Thinking on by default (adaptive); disabling is only valid at effort
+        // high or below (xhigh/max + disabled → 400). Full five-level ladder;
+        // Anthropic's rec: xhigh for coding/agentic, but low/medium punch far
+        // above their weight on this model. 512-token prompt-cache minimum.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        codeExecutionToolType: 'code_execution_20250825',
+        webFetchToolType: 'web_fetch_20260209',
+        // Drop-in successor to Opus 4.8 at identical pricing.
+        inputPricePerMTok: 5,
+        outputPricePerMTok: 25,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 0.5,
+        cacheWritePricePerMTok: 6.25,
+        // Not published at verification time — best-effort estimate (≥ Opus 4.8's).
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'claude-opus-4-8',
+        provider: 'anthropic',
+        label: 'Claude Opus 4.8',
+        description: 'Previous Opus — deep reasoning; the opus-5 refusal fallback',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+        defaultEffortLevel: 'high',
+        // Full five-level ladder (medium was missing here). Anthropic's rec for
+        // coding/agentic on Opus 4.8 is xhigh; the API default is high.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        codeExecutionToolType: 'code_execution_20250825',
+        webFetchToolType: 'web_fetch_20260209',
+        inputPricePerMTok: 5,
+        outputPricePerMTok: 25,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 0.5,
+        cacheWritePricePerMTok: 6.25,
+        knowledgeCutoff: '2026-01-01',
+        // Superseded by claude-opus-5 (same price); still served upstream and the
+        // recommended refusal-fallback target. Selectable under "Older models".
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'claude-sonnet-5',
+        provider: 'anthropic',
+        label: 'Claude Sonnet 5',
+        description: 'Fast & capable — near-Opus coding at Sonnet cost',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh', 'max'],
+        defaultEffortLevel: 'high',
+        // Adaptive thinking on by default; full five-level ladder (medium was
+        // missing here). Default/rec is high, xhigh for the hardest coding/agentic
+        // tasks, medium ≈ Sonnet 4.6 at high.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        codeExecutionToolType: 'code_execution_20250825',
+        webFetchToolType: 'web_fetch_20260209',
+        // Standard pricing. Intro pricing ($2/$10) applies through 2026-08-31 —
+        // billed here at standard so metering never under-charges; revisit after.
+        inputPricePerMTok: 3,
+        outputPricePerMTok: 15,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 0.3,
+        cacheWritePricePerMTok: 3.75,
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'claude-opus-4-6',
+        provider: 'anthropic',
+        label: 'Claude Opus 4.6',
+        description: 'Previous-generation Opus — deep reasoning & complex tasks',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'max'],
+        defaultEffortLevel: 'medium',
+        // No xhigh on the 4.6 family; budget_tokens still accepted but deprecated —
+        // adaptive + effort is the recommended control.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        codeExecutionToolType: 'code_execution_20250825',
+        webFetchToolType: 'web_fetch_20260209',
+        inputPricePerMTok: 5,
+        outputPricePerMTok: 25,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 0.5,
+        cacheWritePricePerMTok: 6.25,
+        knowledgeCutoff: '2025-05-01',
+        // Superseded by claude-opus-4-8; kept selectable (Older models) + priceable.
+        deprecatedAt: '2026-06-16',
+    },
+    {
+        id: 'claude-sonnet-4-6',
+        provider: 'anthropic',
+        label: 'Claude Sonnet 4.6',
+        description: 'Previous-generation Sonnet — fast & balanced',
+        contextWindow: 1_000_000,
+        // Was wrongly 64K — Sonnet 4.6 supports 128K output (models overview).
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 10_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'max'],
+        defaultEffortLevel: 'medium',
+        // No xhigh on the 4.6 family. Anthropic's recommended default for agentic
+        // coding on Sonnet 4.6 is medium (= M); high is its API default (= L).
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20260209',
+        codeExecutionToolType: 'code_execution_20250825',
+        webFetchToolType: 'web_fetch_20260209',
+        inputPricePerMTok: 3,
+        outputPricePerMTok: 15,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 0.3,
+        cacheWritePricePerMTok: 3.75,
+        knowledgeCutoff: '2025-08-01',
+        // Superseded by claude-sonnet-5; kept selectable (Older models) + priceable.
+        deprecatedAt: '2026-07-07',
+    },
+    {
+        id: 'claude-haiku-4-5-20251001',
+        provider: 'anthropic',
+        label: 'Claude Haiku 4.5',
+        description: 'Fastest Anthropic — quick tasks & iteration',
+        contextWindow: 200_000,
+        maxOutputTokens: 64_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        // Haiku 4.5 does NOT support output_config.effort (400) or adaptive
+        // thinking — it stays on the budget_tokens path, so its levels are
+        // budget labels (effortBudgetTokens), not a native effort param.
+        supportedEffortLevels: ['4K', '8K', '16K', '32K'],
+        defaultEffortLevel: '8K',
+        effortBudgetTokens: { '4K': 4000, '8K': 8000, '16K': 16000, '32K': 32000 },
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search_20250305',
+        webFetchToolType: 'web_fetch_20250910',
+        inputPricePerMTok: 1,
+        outputPricePerMTok: 5,
+        // Anthropic 5-minute prompt cache: read 0.1× input, write 1.25× input.
+        cacheReadPricePerMTok: 0.1,
+        cacheWritePricePerMTok: 1.25,
+        knowledgeCutoff: '2025-02-01',
+    },
+    // ---------------------------------------------------------------------------
+    // OpenAI
+    // Verified: https://developers.openai.com/api/docs/pricing (2026-07-28)
+    // GPT-5.6 family GA 2026-07-09: gpt-5.6 is an alias for gpt-5.6-sol; Sol is
+    // the frontier tier, Terra the balanced tier, Luna the cheap/fast tier.
+    // Cached input is billed at 0.1× input. The official pricing page shows no
+    // cache-WRITE premium, but third-party trackers report 1.25× with a 30-min
+    // cache life for the 5.6 family — billed here at 1.25× (conservative;
+    // re-verify against the official docs). reasoning_effort values for 5.6 are
+    // not yet on the docs page — carried over from gpt-5.5 (low|medium|high|
+    // xhigh, default medium); re-verify. Long-context 2× price variants exist
+    // upstream — not modeled (same as the Gemini/Grok >200K tiers).
+    // ---------------------------------------------------------------------------
+    {
+        id: 'gpt-5.6-sol',
+        provider: 'openai',
+        label: 'GPT-5.6 Sol',
+        description: 'OpenAI frontier — the hardest coding & reasoning work',
+        // Reported as 1.05M; floored to 1M until the docs state it (understating
+        // only makes compaction slightly earlier — overstating risks overflow).
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        toolsRequireReasoningOff: true,
+        webSearchToolType: 'web_search',
+        codeExecutionToolType: 'code_interpreter',
+        inputPricePerMTok: 5,
+        outputPricePerMTok: 30,
+        // Cached input 0.1× input; write premium reported 1.25× (see section note).
+        cacheReadPricePerMTok: 0.5,
+        cacheWritePricePerMTok: 6.25,
+        // Not published — best-effort estimate.
+        knowledgeCutoff: '2026-03-01',
+    },
+    {
+        id: 'gpt-5.6-terra',
+        provider: 'openai',
+        label: 'GPT-5.6 Terra',
+        description: 'Balanced OpenAI — everyday coding & tool use',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        toolsRequireReasoningOff: true,
+        webSearchToolType: 'web_search',
+        codeExecutionToolType: 'code_interpreter',
+        // Repriced 2026-07-30 (20% cut from $2.50/$15).
+        inputPricePerMTok: 2,
+        outputPricePerMTok: 12,
+        // Cached input 0.1× input; write premium reported 1.25× (see section note).
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 2.5,
+        // Not published — best-effort estimate.
+        knowledgeCutoff: '2026-03-01',
+    },
+    {
+        id: 'gpt-5.6-luna',
+        provider: 'openai',
+        label: 'GPT-5.6 Luna',
+        description: 'Fast & cheap OpenAI — light tasks & subagents',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        toolsRequireReasoningOff: true,
+        webSearchToolType: 'web_search',
+        codeExecutionToolType: 'code_interpreter',
+        // Repriced 2026-07-30 (80% cut from $1/$6).
+        inputPricePerMTok: 0.2,
+        outputPricePerMTok: 1.2,
+        // Cached input 0.1× input; write premium reported 1.25× (see section note).
+        cacheReadPricePerMTok: 0.02,
+        cacheWritePricePerMTok: 0.25,
+        // Not published — best-effort estimate.
+        knowledgeCutoff: '2026-03-01',
+    },
+    {
+        id: 'gpt-5.5',
+        provider: 'openai',
+        label: 'GPT-5.5',
+        description: 'OpenAI flagship — strong coding & tool use',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        // OpenAI default + agentic-coding rec is medium (= M); 'none' is not
+        // offered (it disables reasoning outright).
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        toolsRequireReasoningOff: true,
+        webSearchToolType: 'web_search',
+        codeExecutionToolType: 'code_interpreter',
+        inputPricePerMTok: 5,
+        outputPricePerMTok: 30,
+        // OpenAI auto-caches at no write premium; cached input billed at 0.1× input.
+        cacheReadPricePerMTok: 0.5,
+        cacheWritePricePerMTok: 5,
+        knowledgeCutoff: '2025-12-01',
+        // Superseded by gpt-5.6-sol (same price); still listed as current by
+        // OpenAI. Selectable under "Older models".
+        deprecatedAt: '2026-07-09',
+    },
+    {
+        id: 'gpt-5.4',
+        provider: 'openai',
+        label: 'GPT-5.4',
+        description: 'Affordable OpenAI frontier — strong coding & tool use',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        toolsRequireReasoningOff: true,
+        webSearchToolType: 'web_search',
+        codeExecutionToolType: 'code_interpreter',
+        inputPricePerMTok: 2.5,
+        outputPricePerMTok: 15,
+        // OpenAI auto-caches at no write premium; cached input billed at 0.1× input.
+        cacheReadPricePerMTok: 0.25,
+        cacheWritePricePerMTok: 2.5,
+        knowledgeCutoff: '2025-08-31',
+        // OpenAI still lists gpt-5.4 as current, but gpt-5.6-terra covers this
+        // tier at the same price — moved to "Older models" (deprecatedAt is OUR
+        // picker taxonomy, not OpenAI's deprecations page).
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'gpt-5.4-mini',
+        provider: 'openai',
+        label: 'GPT-5.4 mini',
+        description: 'Cheap & fast OpenAI — light coding tasks & subagents',
+        contextWindow: 400_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high', 'xhigh'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Tools + ANY reasoning is a 400 on /v1/chat/completions for this family.
+        toolsRequireReasoningOff: true,
+        webSearchToolType: 'web_search',
+        codeExecutionToolType: 'code_interpreter',
+        inputPricePerMTok: 0.75,
+        outputPricePerMTok: 4.5,
+        // OpenAI auto-caches at no write premium; cached input billed at 0.1× input.
+        cacheReadPricePerMTok: 0.075,
+        cacheWritePricePerMTok: 0.75,
+        knowledgeCutoff: '2025-08-31',
+        // OpenAI still lists gpt-5.4-mini as current, but like gpt-5.4 above it's
+        // superseded in our lineup (cheap/fast tier is better served by the newer
+        // models) — moved to "Older models" (deprecatedAt is OUR picker taxonomy,
+        // not OpenAI's deprecations page).
+        deprecatedAt: '2026-08-01',
+    },
+    // ---------------------------------------------------------------------------
+    // Google
+    // Verified: https://ai.google.dev/gemini-api/docs/models
+    //           https://ai.google.dev/gemini-api/docs/pricing
+    //           https://ai.google.dev/gemini-api/docs/thinking
+    // Reasoning control is `thinking_level` (minimal|low|medium|high — cannot be
+    // fully off on 3.x; combining with legacy thinking_budget returns 400).
+    // NOTE: the previous catalog carried a fictional "gemini-3.1-pro" GA id — no
+    // such model exists; the pro tier is (still) gemini-3.1-pro-preview. Safe to
+    // replace outright: the google bond has never been implemented/wired, so no
+    // historical usage can reference the old ids.
+    // ---------------------------------------------------------------------------
+    {
+        id: 'gemini-3.6-flash',
+        provider: 'google',
+        label: 'Gemini 3.6 Flash',
+        description: 'Google agentic flagship — frontier intelligence + grounding',
+        // Window/output not on the pricing page — carried over from 3.5-flash;
+        // re-verify against /docs/models.
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        supportsThinking: true,
+        thinkingBudgetTokens: 10_000,
+        thinkingConfigurable: true,
+        // thinking_level assumed unchanged from 3.5-flash (low|medium|high);
+        // re-verify against /docs/thinking.
+        supportedEffortLevels: ['low', 'medium', 'high'],
+        defaultEffortLevel: 'medium',
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'google_search',
+        codeExecutionToolType: 'code_execution',
+        webFetchToolType: 'url_context',
+        // GA 2026-07-21 — same input price as 3.5-flash, CHEAPER output ($7.50 vs $9).
+        inputPricePerMTok: 1.5,
+        outputPricePerMTok: 7.5,
+        // Gemini context cache: read $0.15/M (0.1× input), no write premium
+        // (storage billed separately per hour — not modeled).
+        cacheReadPricePerMTok: 0.15,
+        cacheWritePricePerMTok: 1.5,
+        // Not published — best-effort estimate.
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'gemini-3.5-flash',
+        provider: 'google',
+        label: 'Gemini 3.5 Flash',
+        description: 'Previous Google agentic flagship — fast, 1M context',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        supportsThinking: true,
+        thinkingBudgetTokens: 10_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high'],
+        defaultEffortLevel: 'medium',
+        // thinking_level: default medium (= M); Google recommends high (= L) for
+        // advanced coding/multi-step planning. 'minimal' exists below low; no
+        // fourth upward tier, so XL is not offered.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'google_search',
+        codeExecutionToolType: 'code_execution',
+        webFetchToolType: 'url_context',
+        inputPricePerMTok: 1.5,
+        outputPricePerMTok: 9,
+        // Gemini context cache: read $0.15/M (0.1× input), no write premium
+        // (storage billed separately per hour — not modeled).
+        cacheReadPricePerMTok: 0.15,
+        cacheWritePricePerMTok: 1.5,
+        knowledgeCutoff: '2025-01-01',
+        // Superseded by gemini-3.6-flash (2026-07-21); still served upstream.
+        // Selectable under "Older models".
+        deprecatedAt: '2026-07-21',
+    },
+    {
+        id: 'gemini-3.1-pro-preview',
+        provider: 'google',
+        label: 'Gemini 3.1 Pro',
+        description: 'Google pro tier — deep reasoning (preview id; no GA id exists)',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        supportsThinking: true,
+        thinkingBudgetTokens: 10_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'medium', 'high'],
+        defaultEffortLevel: 'medium',
+        // thinking_level: low|medium|high only (no minimal). Google's API default
+        // for this model is high (= L); M offers the cost step-down.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'google_search',
+        codeExecutionToolType: 'code_execution',
+        webFetchToolType: 'url_context',
+        // ≤200K-token prompts; >200K bills $4/$18 (tiering not modeled).
+        inputPricePerMTok: 2,
+        outputPricePerMTok: 12,
+        // Gemini context cache: read $0.20/M (≤200K), no write premium (hourly
+        // storage not modeled).
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 2,
+        knowledgeCutoff: '2025-01-01',
+    },
+    // ---------------------------------------------------------------------------
+    // xAI (Grok)
+    // Verified: https://docs.x.ai/developers/models + /developers/grok-4-5
+    //           (2026-07-28)
+    // grok-4.5 (2026-07-08) is the flagship: 500K ctx, reasoning_effort
+    // low|medium|high default high, image input. grok-4.3 stays served as the
+    // value tier with the BIGGER 1M window (reasoning_effort none|low|medium|
+    // high, default low). Max output tokens still not documented for any model.
+    // ---------------------------------------------------------------------------
+    {
+        id: 'grok-4.5',
+        provider: 'xai',
+        label: 'Grok 4.5',
+        description: 'xAI frontier — coding, agentic tasks & knowledge work',
+        contextWindow: 500_000,
+        // Max output not documented by xAI — conservative cap.
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        // reasoning_effort low|medium|high, default high (docs.x.ai/developers/
+        // grok-4-5) — no 'none' tier on 4.5, unlike 4.3.
+        supportedEffortLevels: ['low', 'medium', 'high'],
+        defaultEffortLevel: 'high',
+        // Multimodal input (text + images), text out.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // ≤200K-token prompts; ≥200K bills 2× ($4/$0.60/$12 — tiering not
+        // modeled, same as the Gemini 3.1 Pro >200K tier).
+        inputPricePerMTok: 2,
+        outputPricePerMTok: 6,
+        // xAI cached input billed at a flat $0.30/M, no write premium.
+        cacheReadPricePerMTok: 0.3,
+        cacheWritePricePerMTok: 2,
+        // Official (docs.x.ai): 2026-02-01.
+        knowledgeCutoff: '2026-02-01',
+    },
+    {
+        id: 'grok-4.3',
+        provider: 'xai',
+        label: 'Grok 4.3',
+        description: 'xAI value tier — fast reasoning, bigger 1M context',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['none', 'low', 'medium', 'high'],
+        defaultEffortLevel: 'low',
+        // xAI's default (and its own retirement-routing choice for agentic
+        // workloads) is low (= M); none (= S) disables reasoning for a true fast
+        // mode; step up for hard debugging/architecture turns.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 1.25,
+        outputPricePerMTok: 2.5,
+        // xAI cached input billed at a flat $0.20/M, no write premium.
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 1.25,
+        knowledgeCutoff: '2025-12-01',
+        // Superseded by grok-4.5 as the xAI pick (4.3 keeps the bigger 1M window
+        // — the reason it stays selectable under "Older models").
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'grok-build-0.1',
+        provider: 'xai',
+        label: 'Grok Build',
+        description: 'Agentic coding specialist — fast & cheap (public beta)',
+        contextWindow: 256_000,
+        // Max output not documented by xAI — conservative cap.
+        maxOutputTokens: 64_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Reasoning is always on and NOT configurable (reasoning_effort is not
+        // honored on grok-build) — successor to grok-code-fast-1 (that slug now
+        // auto-routes here).
+        thinkingConfigurable: false,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 1,
+        outputPricePerMTok: 2,
+        // xAI cached input billed at a flat $0.20/M, no write premium.
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 1,
+        // Not published by xAI — best-effort estimate (grok-4-generation base).
+        knowledgeCutoff: '2025-06-01',
+        // Niche coding beta; grok-4.5 is the xAI pick — kept out of the main list.
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'grok-4.20-multi-agent-beta-0309',
+        provider: 'xai',
+        label: 'Grok 4.20',
+        description: 'Older xAI flagship — multi-agent',
+        // Current xAI docs list 1M (the old 2M figure is stale). This beta slug is
+        // an alias of the canonical grok-4.20-multi-agent-0309 — kept under the
+        // original id so saved selections + historical usage stay priceable.
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 16_000,
+        // On the multi-agent model reasoning_effort controls AGENT COUNT, not
+        // reasoning depth — do not drive it from the effort setting.
+        thinkingConfigurable: false,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Repriced by xAI when grok-4.3 launched (was $2/$6).
+        inputPricePerMTok: 1.25,
+        outputPricePerMTok: 2.5,
+        // xAI cached input billed at a flat $0.20/M, no write premium.
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 1.25,
+        knowledgeCutoff: '2024-11-01',
+        deprecatedAt: '2026-04-30',
+        // DISABLED, not deleted: xAI rejects this model on the chat-completions
+        // endpoint outright — "Multi Agent requests are not allowed on chat
+        // completions" (400 on every call, verified live 2026-07-30) — so offering
+        // it in the picker only hands users a model that cannot answer. Their API
+        // also lists the canonical slug as `grok-4.20-multi-agent-0309` (no
+        // `beta-`), with `grok-4.20-0309-reasoning` / `-non-reasoning` as the
+        // chat-completions-capable variants if this family is wanted back.
+        // Stays in the catalogue so saved selections and past usage remain priceable.
+        disabled: true,
+    },
+    {
+        id: 'grok-code-fast-1',
+        provider: 'xai',
+        label: 'Grok Code',
+        description: 'Code specialist — fast & cheap',
+        contextWindow: 256_000,
+        maxOutputTokens: 64_000,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 0.2,
+        outputPricePerMTok: 1.5,
+        // xAI cached input billed at a discount (≈0.25× input), no write premium.
+        cacheReadPricePerMTok: 0.05,
+        cacheWritePricePerMTok: 0.2,
+        knowledgeCutoff: '2024-11-01',
+        // Deprecated by xAI 2026-05-15 (retires 2026-08-15; the slug auto-routes to
+        // grok-build-0.1 until then) — disabled: removed from selection + the
+        // public listing, but getModel() still prices any historical usage. NEVER
+        // delete.
+        disabled: true,
+    },
+    // ---------------------------------------------------------------------------
+    // DeepSeek
+    // Verified: https://api-docs.deepseek.com/quick_start/pricing
+    //           https://api-docs.deepseek.com/guides/thinking_mode
+    //           https://api-docs.deepseek.com/updates/ (2026-07-31)
+    // 2026-07-31: DeepSeek-V4-Flash OFFICIAL API launched in public beta — the
+    // SAME `deepseek-v4-flash` id now serves the re-post-trained 0731 build
+    // (same architecture/size; much stronger agent benchmarks — beats
+    // V4-Pro-Preview on Terminal Bench 2.1 / DeepSWE). No pricing/limit/
+    // capability changes. V4-Pro official release "coming soon" — re-verify
+    // pricing THEN (the announced peak-hour 2× was tied to the V4 official
+    // rollout and is still not on the rate card).
+    // OpenAI/Anthropic-compatible API; text/code only (no vision); 1M context,
+    // 384K max output, automatic context (prompt) caching with ABSOLUTE cache-hit
+    // prices (~1/50–1/120 of miss — not the old 0.1× rule). Launch discount made
+    // PERMANENT 2026-05-23 (Pro $1.74/$3.48 → $0.435/$0.87). Peak-hour 2×
+    // pricing announced for the mid-Jul 2026 "V4 official" release — re-verify
+    // then. Thinking now defaults ENABLED upstream and supports tool calling
+    // (reasoning_effort: high|max), BUT tool loops in thinking mode must replay
+    // assistant reasoning_content on every subsequent request (400 on omission).
+    // The bond explicitly sends thinking:{type:"disabled"} — Synthase runs
+    // DeepSeek as a non-thinking executor (Sonnet plans; DeepSeek executes).
+    // ---------------------------------------------------------------------------
+    {
+        id: 'deepseek-v4-pro',
+        provider: 'deepseek',
+        label: 'DeepSeek V4 Pro',
+        description: 'Frontier-class — rivals top models at low cost',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 384_000,
+        // Run non-thinking (see section note): the executor tool loop would have to
+        // replay reasoning_content across every turn in thinking mode.
+        supportsThinking: false,
+        thinkingBudgetTokens: 0,
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 0.435,
+        outputPricePerMTok: 0.87,
+        // DeepSeek automatic context cache: absolute cache-hit price ($/M).
+        cacheReadPricePerMTok: 0.003625,
+        cacheWritePricePerMTok: 0.435,
+        // Native-China DEFAULT (owner decision 2026-08-01): the US re-host
+        // (DeepInfra) bills ~3× list and ~28× cache reads, and agentic input is
+        // ~94% cache hits, so US processing ran ~5.7× native on real traffic.
+        // Users opt into US per model via the picker's region control.
+        regions: ['cn', 'us'],
+        // The free tier PLANS with this model on the cheap native host (it is the
+        // molecule-dev FREE_TIER_MODELS.plan), so CN is free-tier selectable; the
+        // ~3× US re-host stays paid-only (free users switch to Flash for US).
+        freeTierRegions: ['cn'],
+        // US = DeepInfra, verified 2026-08-01 via api.deepinfra.com/models/
+        // deepseek-ai/DeepSeek-V4-Pro. No cache-write premium (omitted → region
+        // input rate).
+        regionPricing: {
+            us: { inputPricePerMTok: 1.3, outputPricePerMTok: 2.6, cacheReadPricePerMTok: 0.1 },
+        },
+        // The announced peak-hour 2× surcharge (Beijing business hours) is STILL
+        // NOT ACTIVE as of 2026-07-28 — the official rate card lists a single flat
+        // rate, and no switch-over date is published. The pre-wired windows were
+        // REMOVED: they had been over-billing every peak-window turn 2× for weeks
+        // (this is the free-tier default model, so that directly shrank free
+        // users' allowances). Re-add via `peakPricing` the day DeepSeek's rate
+        // card actually shows the surcharge.
+        // Not published by DeepSeek — best-effort estimate.
+        knowledgeCutoff: '2025-07-01',
+    },
+    {
+        id: 'deepseek-v4-flash',
+        provider: 'deepseek',
+        label: 'DeepSeek V4 Flash',
+        description: 'Ultra-cheap & fast — economical agentic coding',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 384_000,
+        // Non-thinking (see section note) — also the right tradeoff for a cheap,
+        // fast tool-calling executor.
+        supportsThinking: false,
+        thinkingBudgetTokens: 0,
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Free-tier default: cheapest model + a fast, non-thinking tool-calling
+        // executor — the model the IDE picks when none is chosen. Exactly one model
+        // in this catalog may carry freeTier (enforced by lookup.test.ts).
+        freeTier: true,
+        inputPricePerMTok: 0.14,
+        outputPricePerMTok: 0.28,
+        // DeepSeek automatic context cache: absolute cache-hit price ($/M).
+        cacheReadPricePerMTok: 0.0028,
+        cacheWritePricePerMTok: 0.14,
+        // Native-China default, matching deepseek-v4-pro (see its note) — even
+        // though Flash's US list price is BELOW native, its cache reads are 6.4×,
+        // and the plan/execute pair defaults to one region deliberately.
+        regions: ['cn', 'us'],
+        // US = DeepInfra, verified 2026-08-01 (api.deepinfra.com/models/…V4-Flash).
+        regionPricing: {
+            us: { inputPricePerMTok: 0.09, outputPricePerMTok: 0.18, cacheReadPricePerMTok: 0.018 },
+        },
+        // Peak-hour surcharge NOT active (see deepseek-v4-pro) — windows removed.
+        // Not published by DeepSeek — best-effort estimate.
+        knowledgeCutoff: '2025-07-01',
+    },
+    // ---------------------------------------------------------------------------
+    // Moonshot (Kimi)
+    // Verified: https://platform.kimi.ai/docs/models +
+    //           /docs/guide/use-kimi-k2-thinking-model (2026-07-28).
+    // The moonshot bond now supports PRESERVED THINKING (ChatMessage.reasoning →
+    // reasoning_content replay through tool loops), which unblocked the two
+    // previously-excluded models: kimi-k3 (2026-07-16 flagship — 2.8T MoE, 1M
+    // ctx, forced thinking, reasoning_effort low|high|max default max upstream)
+    // and kimi-k2.7-code (coding flagship — forced thinking, no depth knob).
+    // kimi-k2.x thinking stays on/off only; the bond disables it for those by
+    // default (KIMI_REASONING_EFFORT env tunes it).
+    // ---------------------------------------------------------------------------
+    {
+        id: 'kimi-k3',
+        provider: 'moonshot',
+        label: 'Kimi K3',
+        description: 'Moonshot flagship — 2.8T open weights, 1M context, multimodal',
+        contextWindow: 1_000_000,
+        // Max output not documented — conservative cap (matches the K2.x family).
+        maxOutputTokens: 65_535,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking is FORCED ON; depth rides reasoning_effort (low|high|max). The
+        // upstream default is max — we default to high so agentic loops aren't
+        // pinned to the slowest/most expensive tier unless the user asks for it.
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['low', 'high', 'max'],
+        defaultEffortLevel: 'high',
+        // Text + image + video input.
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 3,
+        outputPricePerMTok: 15,
+        // Automatic context cache: absolute cache-hit price ($0.30/M = 0.1× input).
+        cacheReadPricePerMTok: 0.3,
+        cacheWritePricePerMTok: 3,
+        // No US re-host exists (not on DeepInfra) — pinned to native China.
+        regions: ['cn'],
+        // Not published — best-effort estimate.
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'kimi-k2.7-code',
+        provider: 'moonshot',
+        label: 'Kimi K2.7 Code',
+        description: 'Moonshot coding specialist — token-efficient agentic coding',
+        contextWindow: 262_144,
+        maxOutputTokens: 65_535,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking forced on, NO depth knob (reasoning_effort unsupported here) —
+        // the bond sends neither param and replays reasoning_content.
+        thinkingConfigurable: false,
+        // Coding model — vision not documented; conservative.
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 0.95,
+        outputPricePerMTok: 4,
+        // Automatic context cache: absolute cache-hit price ($0.19/M = 0.2× input).
+        cacheReadPricePerMTok: 0.19,
+        cacheWritePricePerMTok: 0.95,
+        // US default (DeepInfra bills below native here). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.74, outputPricePerMTok: 3.5, cacheReadPricePerMTok: 0.15 },
+        },
+        // Not published — best-effort estimate.
+        knowledgeCutoff: '2025-10-01',
+        // kimi-k3 is the Moonshot pick; the coding specialist stays selectable
+        // under "Older models" for anyone who wants the cheaper tier.
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'kimi-k2.6',
+        provider: 'moonshot',
+        label: 'Kimi K2.6',
+        description: 'Multimodal agent — strong sequential tool use',
+        contextWindow: 262_144,
+        // Output shares the 256K context window (max_tokens default 32K upstream);
+        // practical cap.
+        maxOutputTokens: 65_535,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking is on/off only (default on upstream; the bond disables it by
+        // default for the executor loop — tune via KIMI_REASONING_EFFORT).
+        thinkingConfigurable: false,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Native platform.kimi.ai pricing (the old $0.68/$3.41 was OpenRouter's
+        // blended third-party rate).
+        inputPricePerMTok: 0.95,
+        outputPricePerMTok: 4,
+        cacheReadPricePerMTok: 0.16,
+        cacheWritePricePerMTok: 0.95,
+        // US default (DeepInfra bills below native here). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.75, outputPricePerMTok: 3.5, cacheReadPricePerMTok: 0.15 },
+        },
+        knowledgeCutoff: '2025-04-01',
+        // Superseded by kimi-k3; moved to "Older models".
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'kimi-k2.5',
+        provider: 'moonshot',
+        label: 'Kimi K2.5',
+        description: 'Older Kimi — strong sequential tool use',
+        contextWindow: 262_144,
+        maxOutputTokens: 65_535,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking on/off only; no preserved-thinking support upstream.
+        thinkingConfigurable: false,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Native platform.kimi.ai pricing.
+        inputPricePerMTok: 0.6,
+        outputPricePerMTok: 3,
+        cacheReadPricePerMTok: 0.1,
+        cacheWritePricePerMTok: 0.6,
+        // US default (DeepInfra bills below native here). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.45, outputPricePerMTok: 2.25, cacheReadPricePerMTok: 0.07 },
+        },
+        knowledgeCutoff: '2024-04-01',
+        // Superseded by kimi-k2.6 (still served upstream, no announced retirement);
+        // kept selectable (Older models) + priceable.
+        deprecatedAt: '2026-04-01',
+    },
+    // ---------------------------------------------------------------------------
+    // MiniMax
+    // Verified: https://platform.minimax.io/docs/guides/models-intro
+    //           https://platform.minimax.io/docs/guides/pricing-paygo.md
+    // minimax-m3 (2026-06-01) is the flagship: 1M context (≥512K guaranteed),
+    // multimodal (text+image+video in), thinking adaptive|disabled (no budget /
+    // effort levels). Run non-thinking like the DeepSeek executors — with
+    // thinking on, reasoning must be carried through tool loops. m2.7 thinking
+    // CANNOT be disabled upstream.
+    // ---------------------------------------------------------------------------
+    {
+        id: 'minimax-m3',
+        provider: 'minimax',
+        label: 'MiniMax M3',
+        description: 'Agentic flagship — 1M context, multimodal, great value',
+        contextWindow: 1_048_576,
+        // Upstream recommends 128K max_tokens (hard max 512K).
+        maxOutputTokens: 131_072,
+        // Run non-thinking (the bond sends thinking:{type:"disabled"} for M3);
+        // native control is adaptive|disabled only — no depth levels.
+        supportsThinking: false,
+        thinkingBudgetTokens: 0,
+        thinkingConfigurable: false,
+        supportsVision: true,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // ≤512K-token prompts; >512K bills 2× (tiering not modeled).
+        inputPricePerMTok: 0.3,
+        outputPricePerMTok: 1.2,
+        cacheReadPricePerMTok: 0.06,
+        // M3 cache-write price not published — M2.7's rate (≥ input as required).
+        cacheWritePricePerMTok: 0.375,
+        // US default. DeepInfra list matches native; only the cache write differs
+        // (no premium → region input rate). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.3, outputPricePerMTok: 1.2, cacheReadPricePerMTok: 0.06 },
+        },
+        // From the official HF chat template ("Knowledge cutoff: January 2026").
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'minimax-m2.7',
+        provider: 'minimax',
+        label: 'MiniMax M2.7',
+        description: 'Agentic productivity — strong value for coding',
+        contextWindow: 204_800,
+        maxOutputTokens: 196_608,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking is ALWAYS ON for M2.x (cannot be disabled) — no depth control.
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // Native platform repriced (was $0.25/$1.00).
+        inputPricePerMTok: 0.3,
+        outputPricePerMTok: 1.2,
+        cacheReadPricePerMTok: 0.06,
+        cacheWritePricePerMTok: 0.375,
+        // US default (DeepInfra still bills the pre-reprice rate). Verified
+        // 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.25, outputPricePerMTok: 1, cacheReadPricePerMTok: 0.05 },
+        },
+        knowledgeCutoff: '2025-09-01',
+        // Superseded by minimax-m3 (same price, 1M ctx, multimodal); moved to
+        // "Older models".
+        deprecatedAt: '2026-07-28',
+    },
+    {
+        id: 'minimax-m2.5',
+        provider: 'minimax',
+        label: 'MiniMax M2.5',
+        description: 'Older MiniMax — strong value for coding',
+        contextWindow: 196_608,
+        maxOutputTokens: 196_608,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking always on for M2.x — no depth control.
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 0.3,
+        outputPricePerMTok: 1.2,
+        cacheReadPricePerMTok: 0.03,
+        cacheWritePricePerMTok: 0.375,
+        // No US re-host exists (not on DeepInfra) — pinned to native China.
+        regions: ['cn'],
+        knowledgeCutoff: '2025-01-01',
+        // Superseded by minimax-m3 (legacy upstream, still served); kept selectable
+        // (Older models) + priceable.
+        deprecatedAt: '2026-03-18',
+    },
+    // ---------------------------------------------------------------------------
+    // Alibaba (Qwen)
+    // Verified: https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+    //           https://www.alibabacloud.com/help/en/model-studio/qwen-coder
+    //           https://openrouter.ai/qwen/qwen3.7-max
+    // qwen3.7-max (2026-05-21) is the agentic flagship — Alibaba's own Qwen-Coder
+    // docs now recommend the general-purpose models over Qwen-Coder. Its thinking
+    // uses enable_thinking (default ON for the 3.7 series) + thinking_budget
+    // (token cap) — a real budget param, so effort scales the budget. The
+    // qwen3-coder models are NON-thinking (previous catalog entry was wrong).
+    // Prices are DashScope international list rates (the bond calls DashScope,
+    // not OpenRouter; a 50%-off promo currently applies — billed at list).
+    // ---------------------------------------------------------------------------
+    {
+        id: 'qwen3.7-max',
+        provider: 'alibaba',
+        label: 'Qwen3.7 Max',
+        description: 'Alibaba agentic flagship — 1M context, hybrid thinking',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 65_536,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // enable_thinking + thinking_budget: a controllable token budget → effort
+        // scales the budget (no native level names).
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['4K', '8K', '16K', '32K'],
+        defaultEffortLevel: '8K',
+        effortBudgetTokens: { '4K': 4000, '8K': 8000, '16K': 16000, '32K': 32000 },
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        inputPricePerMTok: 2.5,
+        outputPricePerMTok: 7.5,
+        // Implicit context cache: read ≈0.2× input, no write premium.
+        cacheReadPricePerMTok: 0.5,
+        cacheWritePricePerMTok: 2.5,
+        // US default. DeepInfra bills identical rates (no regionPricing needed).
+        // Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        // Not published by Alibaba — best-effort estimate.
+        knowledgeCutoff: '2026-01-01',
+    },
+    {
+        id: 'qwen3-coder-plus',
+        provider: 'alibaba',
+        label: 'Qwen3 Coder Plus',
+        description: 'Coding specialist — 1M context',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 65_536,
+        // Qwen3-Coder models support ONLY non-thinking mode (no thinking control
+        // at all) — the previous "configurable thinking" entry was wrong.
+        supportsThinking: false,
+        thinkingBudgetTokens: 0,
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        // DashScope international list rate, flat across input tiers (the old
+        // $0.65/$3.25 was OpenRouter's promo rate).
+        inputPricePerMTok: 1,
+        outputPricePerMTok: 5,
+        // Implicit context cache: read 0.2× input, no write premium.
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 1,
+        // US default (DeepInfra bills well below native). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.3, outputPricePerMTok: 1, cacheReadPricePerMTok: 0.1 },
+        },
+        knowledgeCutoff: '2025-06-01',
+        // Alibaba itself recommends the general-purpose models over Qwen-Coder;
+        // qwen3.7-max is the pick — moved to "Older models".
+        deprecatedAt: '2026-07-28',
+    },
+    // ---------------------------------------------------------------------------
+    // Zhipu (GLM)
+    // Verified: https://docs.z.ai/guides/overview/pricing
+    //           https://docs.z.ai/api-reference/llm/chat-completion
+    // glm-5.2 (standalone API since 2026-06-16) is the flagship. It is the ONLY
+    // GLM model with reasoning_effort (values minimal|none|low|medium|high|
+    // xhigh|max; low/medium coerce to high, xhigh coerces to max — effective
+    // levels are high|max plus minimal/none = skip thinking; default max).
+    // glm-5 has thinking on/off only and was REPRICED (was $0.72/$2.30).
+    // ---------------------------------------------------------------------------
+    {
+        id: 'glm-5.2',
+        provider: 'zhipu',
+        label: 'GLM-5.2',
+        description: 'Open-source SOTA agentic — 1M context',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 131_072,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        thinkingConfigurable: true,
+        supportedEffortLevels: ['minimal', 'high', 'max'],
+        defaultEffortLevel: 'high',
+        // Z.ai's default is max (= L, positioned for long-horizon coding); M =
+        // high is the balanced tier; minimal (= S) skips thinking for fast turns.
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search',
+        inputPricePerMTok: 1.4,
+        outputPricePerMTok: 4.4,
+        // GLM context cache: read ≈0.19× input, no write premium.
+        cacheReadPricePerMTok: 0.26,
+        cacheWritePricePerMTok: 1.4,
+        // US default (DeepInfra bills ~half native). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.75, outputPricePerMTok: 2.4, cacheReadPricePerMTok: 0.14 },
+        },
+        // Not published by Z.ai — best-effort estimate.
+        knowledgeCutoff: '2025-06-01',
+    },
+    {
+        id: 'glm-5',
+        provider: 'zhipu',
+        label: 'GLM-5',
+        description: '77.8% SWE-bench — open-source agentic, 200K context',
+        contextWindow: 202_752,
+        maxOutputTokens: 131_072,
+        supportsThinking: true,
+        thinkingBudgetTokens: 8_000,
+        // Thinking on/off only (reasoning_effort is glm-5.2-only) — no depth
+        // control; defaults to enabled upstream.
+        thinkingConfigurable: false,
+        supportsVision: false,
+        supportsPromptCaching: true,
+        supportsTools: true,
+        webSearchToolType: 'web_search',
+        // Repriced by Z.ai around the GLM-5.1 launch (was $0.72/$2.30).
+        inputPricePerMTok: 1,
+        outputPricePerMTok: 3.2,
+        // GLM context cache: read 0.2× input, no write premium.
+        cacheReadPricePerMTok: 0.2,
+        cacheWritePricePerMTok: 1,
+        // US default (DeepInfra bills below native). Verified 2026-08-01.
+        regions: ['us', 'cn'],
+        regionPricing: {
+            us: { inputPricePerMTok: 0.6, outputPricePerMTok: 2.08, cacheReadPricePerMTok: 0.12 },
+        },
+        knowledgeCutoff: '2025-01-01',
+        // Superseded by glm-5.2; moved to "Older models".
+        deprecatedAt: '2026-07-28',
+    },
+];