feat(ai-gemini): add Nano Banana 2.1 image model #1640

Add Nano Banana 2.1 (gemini-nano-banana-2.1) to @tanstack/ai-gemini as a native image model.

Why the daily sync did not add it

OpenRouter listed google/gemini-nano-banana-2.1 on 2026-10-06. The Sync Model Metadata run that day added it to @tanstack/ai-openrouter and @tanstack/ai-vercel-gateway (#1636), but not to @tanstack/ai-gemini.

That is by design. scripts/sync-provider-models.ts drops every new model whose output modalities include image (outputsImage), because the Gemini image arrays need hand-written entries: the size and provider-option type maps and the native image union. So each new native image model needs a manual registration, as #870 did for Nano Banana 2 Lite.

Verified facts

From GET https://generativelanguage.googleapis.com/v1beta/models on 2026-10-07:

• Model id: gemini-nano-banana-2.1 (display name "Nano Banana 2.1"). The native id matches the OpenRouter id. It does not follow the gemini-<version>-flash-image pattern of the earlier models.
• Methods: generateContent, countTokens, batchGenerateContent
• Token limits: 65536 input, 65536 output (same as gemini-3.1-flash-image)
• Modalities (OpenRouter): output image + text

Not yet verified

• Supported aspect ratios and resolution tiers (1K / 2K / 4K)
• Pricing
• Whether it accepts image input in the same way as gemini-3.1-flash-image

Confirm these against Google's model page or a live call before writing the type maps.

Implementation

Follow #870:

• packages/ai-gemini/src/model-meta.ts: add the model constant and register it in GEMINI_IMAGE_MODELS
• packages/ai-gemini/src/image/image-provider-options.ts: add it to the GeminiNativeImageModels union (sizes and image input modalities)
• Unit test for native generateContent routing and size parsing
• E2E coverage as the repo requires
• Changeset, docs (docs/adapters/gemini.md, docs/media/image-generation.md), and the media-generation skill model lists

Follow-up, out of scope here

Image models still need a manual issue each time. Generating the Gemini image entries from a catalog that carries native ids and image capabilities would remove that step; #1308 (native model sync from modelschemas) is the likely place.