Skip to main content
Use the OpenAI or Anthropic dialect supported by your client. The gateway translates requests onto a provider protocol that can serve the selected model and operation. A route’s existence does not establish support for every provider. Check the model’s modality and test the request shape your application uses.

​
Model names

Send the exact qualified model id, such as openai-prod/<model-name>, returned by the catalog. The provider prefix selects the configured connection. A bare upstream name can fail to resolve or become ambiguous when more than one connection serves it. Disabled or archived models cannot route. Disabled or archived providers also prevent their models from routing. Catalog queries can include models behind disabled providers when requested, with is_routable: false.

​
Tools and conversation state

Tool definitions and tool results travel in the client’s request dialect. Your application executes client-side tools and sends their results in a later request. Hosted tools such as provider-run web search are unavailable through the Responses compatibility surface. The gateway does not provide the Assistants API, vector stores, batches, fine-tuning, or Anthropic token counting. Responses requests must carry the conversation state they need. Anthropic metadata is accepted and ignored; unsupported conversation-editing context_management is refused.

​
Provider-specific requests

Reasoning effort, structured output, images, audio, cache controls, and provider-specific parameters depend on the model family and provider. A parameter your client accepts may still be refused for the selected model. Use the returned error’s field and reason to adjust the request. For Vertex image generation, Gemini image models use Gemini’s native image API. Responses return base64 image data in PNG or JPEG. Unsupported size ratios, URL responses, and transparent backgrounds are refused. See the image constraints before using OpenAI image options with Vertex. See Model Gateway errors for resolution, compatibility, throttling, and upstream errors.