Google AI Studio

Configure a deployment

Send this body to POST /admin/deployments with an operator credential. Replace the example credentials and choose a model available to your provider account. See Creating deployments for the shared configuration fields.

{ "publicModel": "gemini", "adapterKey": "googleaistudio", "upstreamModel": "gemini-3-pro", "credentials": { "apiKey": "..." } }

Behavior and limitations

• Credentials: apiKey only.
• Default transport: generateContent for text and image operations, generate_videos for Veo video generation, and embedContent/batchEmbedContents for embeddings. Gemini Omni Flash must override video.generate to interactions.
• Tool schemas are translated, not passed through. Non-strict tools and pre-Gemini-3 models use Gemini's OpenAPI-shaped parameters field and its narrower schema translator. Gemini 3 and later models declare capabilities.strictTools: true; a request containing an OpenAI strict: true function uses parametersJsonSchema and Gemini's schema-constrained VALIDATED mode. Pre-Gemini-3 models reject strict requests instead of silently downgrading them. This is generation-time enforcement; the gateway does not coerce arguments or retry a hidden repair request. Constructs outside Google's supported schema subset may still require a simpler tool definition.
• Reasoning: gemini_level (discrete thinkingLevel, with xhigh mapped to high) or gemini_budget (raw thinkingBudget tokens), per model — see Reasoning.
• Thought signatures round-trip statelessly inside the tool call id. When Gemini returns a thoughtSignature on a function call, the gateway embeds it in the public tool call id as <id>__thought__<signature> (LiteLLM-compatible) across all three surfaces. Any client that echoes tool call ids verbatim — every standard OpenAI/Anthropic client, including the Vercel AI SDK — round-trips it with no changes and no server-side state. The signature is also mirrored as provider_specific_fields.thought_signature and extra_content.google.thought_signature on the rendered tool call, and all three inbound forms are accepted. See Provider-specific fields.
• Embeddings accept text input and float vectors only through this gateway — no tokenized input, no base64 encoding — regardless of what the catalog profile might otherwise allow.
• Image quality maps to native thinking on the models that expose the control (Gemini 3.1 Flash Image and Flash Lite Image), which declare the rungs low (thinkingLevel: minimal) and high (thinkingLevel: high); other rungs snap onto those two. auto or an omitted quality sends no thinking level and Gemini's own default applies. Models without the control, such as Gemini 3 Pro Image, declare no rungs at all. See Images.
• Veo uses the Gemini :predictLongRunning REST flow. Image/video references must be data URLs because the REST request sends them as inlineData; multiple image references are forwarded as referenceImages, while video_url is used for Veo extension on models that declare it.
• Gemini Omni Flash 1.1 uses the Interactions API for text-to-video, image-to-video, reference-to-video, edit, extend, and frame interpolation. The gateway runs interactions in the background, polls by interaction id, and fetches inline output into object storage without storing base64 in Postgres. Files API inputs, multi-turn previous_interaction_id, audio references, and URI-only 1080p/4K output are intentionally unsupported.
• Catalog: src/adapters/google/catalog.json (folder is google; adapter key is googleaistudio — the one place today those two names don't match).

Next steps

• Reasoning — gemini_level / gemini_budget in detail.
• Embeddings — the reduced Google embedding contract.
• Images — the quality-to-thinking mapping.
• Videos — the async video lifecycle and object storage.