catalog: toggle-only reasoning models advertise a fake graded thinking ladder (missing thinkingLevelMap)

[code-yeongyu]

Summary

A bounded catalog-accuracy audit (run as part of the per-model reasoning work, plan fast-reasoning-effort-commands) found that 62 builtin catalog models advertise a graded reasoning ladder they do not actually support.

getSupportedThinkingLevels() (packages/ai/src/models.ts:917-928) returns the FULL extended ladder for any reasoning: true model that has no thinkingLevelMap. Toggle-only models (thinkingFormat zai, qwen, qwen-chat-template, string-thinking, deepseek, together) therefore expose fake minimal/low/medium/xhigh levels, and the new classifyReasoningCapability() helper labels them graded instead of on-off/always-on.

Why this is filed instead of fixed inline

The fix must go through the generator (packages/ai/scripts/generate-models.ts) + npm run hydrate:model-data — never hand-edits to generated files. A scratch hydrate on 2026-08-16 showed regeneration currently drags in unrelated upstream catalog drift:

• new models: glm-5.3, deepseek-v4-pro-0813, @cf/deepseek-v4-flash-0731, gemini-3.7-flash, Qwen2.5-Coder-32B-Instruct, meta/muse-glimmer-30b
• removal: cloudflare-ai-gateway: claude-3-5-haiku
• limit bump: baseten 1000000 -> 1048576

Landing that churn inside a behavioral PR would make it unreviewable, so the audit was deliberately split out.

Recommended sequence

1. A standalone catalog refresh PR that lands the upstream drift alone (pure hydrate:model-data output, reviewed as data).
2. Then a generator-source PR adding the missing thinkingLevelMap entries (sections: zai, moonshot, xiaomi, qwen-token-plan, alibaba-token-plan, opencode*) plus packages/coding-agent/test/catalog-thinking-map.test.ts asserting each corrected model's capability.

Models flagged needs-map (62)

| alibaba-token-plan | deepseek-v3.2 | deepseek | none | models.dev alibaba-token-plan: reasoning_options=[{type:'toggle'}]; repo lists deepseek-v3.2 in QWEN_TOKEN_PLAN_REASONING_EFFORT_UNSUPPORTED_MODEL_IDS | needs-map (toggle) |
| alibaba-token-plan | kimi-k2.5 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| alibaba-token-plan | kimi-k2.6 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| alibaba-token-plan | kimi-k2.7-code | deepseek | none | Moonshot doc: kimi-k2.7-code thinking always-on (moonshotapi). models.dev claims toggle on DashScope/alibaba endpoints | needs-map (toggle on resellers; moonshot-direct is always-on and already mapped) |
| alibaba-token-plan | qwen3.6-flash | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| alibaba-token-plan | qwen3.6-plus | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| alibaba-token-plan | qwen3.7-max | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| alibaba-token-plan | qwen3.7-plus | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| alibaba-token-plan | qwen3.8-max | qwen | none | QwenCloud thinking.md: qwen3.8-max reasoning_effort options low/medium/xhigh (default xhigh) | needs-map (graded low/medium/xhigh) |
| alibaba-token-plan | qwen3.8-max-preview | qwen | none | QwenCloud thinking.md: qwen3.8-max reasoning_effort options low/medium/xhigh (default xhigh) | needs-map (graded low/medium/xhigh) |
| moonshotai | kimi-k2-thinking | deepseek | none | Moonshot models.md: 'kimi-k2 series discontinued on May 25, 2026'; no current per-parameter doc; family-consistent toggle | needs-map (toggle; undocumented-primary — model deprecated) |
| moonshotai | kimi-k2-thinking-turbo | deepseek | none | Moonshot models.md: 'kimi-k2 series discontinued on May 25, 2026'; family-consistent toggle | needs-map (toggle; undocumented-primary — model deprecated) |
| moonshotai | kimi-k2.5 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| moonshotai | kimi-k2.6 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| moonshotai-cn | kimi-k2-thinking | deepseek | none | Moonshot models.md: 'kimi-k2 series discontinued on May 25, 2026'; no current per-parameter doc; family-consistent toggle | needs-map (toggle; undocumented-primary — model deprecated) |
| moonshotai-cn | kimi-k2-thinking-turbo | deepseek | none | Moonshot models.md: 'kimi-k2 series discontinued on May 25, 2026'; family-consistent toggle | needs-map (toggle; undocumented-primary — model deprecated) |
| moonshotai-cn | kimi-k2.5 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| moonshotai-cn | kimi-k2.6 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| opencode | kimi-k2.6 | deepseek | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| opencode-go | qwen3.6-plus | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| qwen-token-plan | MiniMax-M2.5 | qwen | none | models.dev reasoning_options=[] (unclassified); repo already maps together MiniMax family as toggle {minimal:null,low:null,medium:null} | needs-map (toggle; unclassified-secondary) |
| qwen-token-plan | deepseek-v3.2 | qwen | none | models.dev alibaba-token-plan: reasoning_options=[{type:'toggle'}]; repo lists deepseek-v3.2 in QWEN_TOKEN_PLAN_REASONING_EFFORT_UNSUPPORTED_MODEL_IDS | needs-map (toggle) |
| qwen-token-plan | kimi-k2.5 | qwen | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| qwen-token-plan | kimi-k2.6 | qwen | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| qwen-token-plan | kimi-k2.7-code | qwen | none | Moonshot doc: kimi-k2.7-code thinking always-on (moonshotapi). models.dev claims toggle on DashScope/alibaba endpoints | needs-map (toggle on resellers; moonshot-direct is always-on and already mapped) |
| qwen-token-plan | qwen3.6-flash | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| qwen-token-plan | qwen3.6-plus | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| qwen-token-plan | qwen3.7-max | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| qwen-token-plan | qwen3.7-plus | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| qwen-token-plan-cn | MiniMax-M2.5 | qwen | none | models.dev reasoning_options=[] (unclassified); repo already maps together MiniMax family as toggle {minimal:null,low:null,medium:null} | needs-map (toggle; unclassified-secondary) |
| qwen-token-plan-cn | deepseek-v3.2 | qwen | none | models.dev alibaba-token-plan: reasoning_options=[{type:'toggle'}]; repo lists deepseek-v3.2 in QWEN_TOKEN_PLAN_REASONING_EFFORT_UNSUPPORTED_MODEL_IDS | needs-map (toggle) |
| qwen-token-plan-cn | kimi-k2.5 | qwen | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| qwen-token-plan-cn | kimi-k2.6 | qwen | none | Moonshot Thinking Models doc: thinking.type enabled/disabled; reasoning_effort 'Not supported' | needs-map (toggle) |
| qwen-token-plan-cn | kimi-k2.7-code | qwen | none | Moonshot doc: kimi-k2.7-code thinking always-on (moonshotapi). models.dev claims toggle on DashScope/alibaba endpoints | needs-map (toggle on resellers; moonshot-direct is always-on and already mapped) |
| qwen-token-plan-cn | qwen3.6-flash | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| qwen-token-plan-cn | qwen3.6-plus | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| qwen-token-plan-cn | qwen3.7-max | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| qwen-token-plan-cn | qwen3.7-plus | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| qwen-token-plan-individual | qwen3.6-flash | qwen | none | QwenCloud thinking.md: 'Hybrid: toggle thinking on or off per request with enable_thinking'; reasoning_effort not documented for 3.6 (only 3.8-max) | needs-map (toggle) |
| qwen-token-plan-individual | qwen3.7-max | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| qwen-token-plan-individual | qwen3.7-plus | qwen | none | QwenCloud thinking.md: enable_thinking hybrid; thinking_budget supported but reasoning_effort not documented for 3.7 | needs-map (toggle) |
| xiaomi | mimo-v2-flash | deepseek | none | Xiaomi OpenAI-compat API doc: thinking.type enabled/disabled (named for v2.5); MiMo-V2 series deprecated June 30; no reasoning_effort anywhere in doc | needs-map (toggle; series deprecated) |
| xiaomi | mimo-v2-omni | deepseek | none | Xiaomi OpenAI-compat API doc: thinking.type enabled/disabled (named for v2.5); MiMo-V2 series deprecated June 30; no reasoning_effort anywhere in doc | needs-map (toggle; series deprecated) |
| xiaomi | mimo-v2-pro | deepseek | none | Xiaomi OpenAI-compat API doc: thinking.type enabled/disabled (named for v2.5); MiMo-V2 series deprecated June 30; no reasoning_effort anywhere in doc | needs-map (toggle; series deprecated) |
| xiaomi | mimo-v2.5 | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled'; no reasoning_effort field in doc | needs-map (toggle) |
| xiaomi | mimo-v2.5-pro | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled' | needs-map (toggle) |
| xiaomi | mimo-v2.5-pro-ultraspeed | deepseek | none | Xiaomi OpenAI-compat API doc (v2.5 family toggle); ultraspeed variant doc'd via marketing page; no graded effort | needs-map (toggle) |
| xiaomi-token-plan-ams | mimo-v2-pro | deepseek | none | Xiaomi OpenAI-compat API doc: thinking.type enabled/disabled (named for v2.5); MiMo-V2 series deprecated June 30; no reasoning_effort anywhere in doc | needs-map (toggle; series deprecated) |
| xiaomi-token-plan-ams | mimo-v2.5 | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled'; no reasoning_effort field in doc | needs-map (toggle) |
| xiaomi-token-plan-ams | mimo-v2.5-pro | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled' | needs-map (toggle) |
| xiaomi-token-plan-cn | mimo-v2-pro | deepseek | none | Xiaomi OpenAI-compat API doc: thinking.type enabled/disabled (named for v2.5); MiMo-V2 series deprecated June 30; no reasoning_effort anywhere in doc | needs-map (toggle; series deprecated) |
| xiaomi-token-plan-cn | mimo-v2.5 | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled'; no reasoning_effort field in doc | needs-map (toggle) |
| xiaomi-token-plan-cn | mimo-v2.5-pro | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled' | needs-map (toggle) |
| xiaomi-token-plan-sgp | mimo-v2-pro | deepseek | none | Xiaomi OpenAI-compat API doc: thinking.type enabled/disabled (named for v2.5); MiMo-V2 series deprecated June 30; no reasoning_effort anywhere in doc | needs-map (toggle; series deprecated) |
| xiaomi-token-plan-sgp | mimo-v2.5 | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled'; no reasoning_effort field in doc | needs-map (toggle) |
| xiaomi-token-plan-sgp | mimo-v2.5-pro | deepseek | none | Xiaomi OpenAI-compat API doc: 'mimo-v2.5-pro, mimo-v2.5: default enabled; Available options: enabled, disabled' | needs-map (toggle) |
| zai | glm-4.7 | zai | none | Z.AI API ref: reasoning_effort is 'Only supported by GLM-5.2'; GLM-4.7 accepts thinking.type enabled/disabled only | needs-map (toggle) |
| zai | glm-5-turbo | zai | none | Z.AI API ref: reasoning_effort 'Only supported by GLM-5.2'; thinking.type enabled/disabled | needs-map (toggle) |
| zai | glm-5.2-highspeed | zai | none | models.dev zai-coding-plan: reasoning_options=[{type:'effort',values:['high','max']}] (same as glm-5.2); id absent from Z.AI public API model enum | needs-map (graded high/max; undocumented-primary) |
| zai-coding-cn | glm-4.7 | zai | none | Z.AI API ref: reasoning_effort is 'Only supported by GLM-5.2'; GLM-4.7 accepts thinking.type enabled/disabled only | needs-map (toggle) |
| zai-coding-cn | glm-5-turbo | zai | none | Z.AI API ref: reasoning_effort 'Only supported by GLM-5.2'; thinking.type enabled/disabled | needs-map (toggle) |
| zai-coding-cn | glm-5.2-highspeed | zai | none | models.dev zai-coding-plan: reasoning_options=[{type:'effort',values:['high','max']}] (same as glm-5.2); id absent from Z.AI public API model enum | needs-map (graded high/max; undocumented-primary) |

Additional findings

• Xiaomi transport contradiction: compat flag supportsDisabledThinking: false contradicts Xiaomi's own API docs (thinking.type: "disabled" is accepted). A toggle map exposing off would advertise a level the current transport cannot actually disable. Transport-side fix, separate concern.
• zai/glm-5.2: map omits xhigh although Z.AI docs map xhigh -> max. Outside the "no map" filter of this audit, but wrong.
• Deprecated entries: kimi-k2-thinking, kimi-k2-thinking-turbo (Moonshot discontinued the kimi-k2 series 2026-05-25) and MiMo-V2 are better deleted than mapped.
• Weak-source rows: MiniMax-M2.5 (models.dev reasoning_options=[]) and glm-5.2-highspeed (absent from Z.AI's public model enum) rest on secondary sources; confirm before mapping.

Provider documentation quotes with URLs are recorded in the audit artifact produced by that run.

Filed by the fast-reasoning-effort-commands work plan; the behavioral PR deliberately leaves packages/ai byte-identical (84 generated artifacts verified unchanged).