Problem#
TanStack AI wants to fill its native model lists from GET /v1/models?provider=… in place of the OpenRouter catalogue. For OpenAI, Anthropic and Grok the data is good enough. For Gemini, two gaps would add wrong models.
- Non-chat models have
activity: "chat". Examples:
lyria-3.5, lyria-3-clip-preview: music (output includes audio).
gemini-nano-banana-2.1, nano-banana-pro-preview: image generation.
gemini-omni-1.1-flash, gemini-omni-1.2-flash, gemini-omni-flash-preview: video. modalities is null.
gemini-3.5-transcribe, gemini-3.5-transcribe-live: transcription.
gemini-3.8-live, gemini-3.1-flash-live-preview and the other -live models: Live API only, not generateContent.
- 26 chat models have no
releasedAt, for example gemini-3.8-flash, gemini-3.7-flash, gemini-4-argon, gemini-4-carbon. The sync uses the release date to skip old models.
Also, 17 Gemini chat models and 3 OpenAI chat models (gpt-4o-2024-11-20, gpt-5-search-api-2025-10-14, gpt-6.1-terra) have pricing: null.
Ask#
- Set
activity from what the model does (music, image, video, transcription, realtime), not from the endpoint it shares.
- Fill
releasedAt (or firstSeenAt, if the release date is unknown) for every model.
- Fill pricing where Google and OpenAI publish it.
Acceptance#
GET /v1/models?provider=gemini with activity: "chat" lists only models that generateContent serves as chat.
- Every model has
releasedAt or firstSeenAt.
Context: TanStack/ai#2104 (moving scripts/sync-provider-models.ts from OpenRouter to modelschemas).