Skip to main content
The gateway translates supported request formats to the selected model’s serving provider. Use the format your application already speaks, and choose a model that supports the task. Compatibility applies to the documented fields and endpoints, not every feature in a provider’s SDK.

​
Text and conversations

For Gemini-native streaming, use :streamGenerateContent. For the other text endpoints, set stream: true. Use full creator-prefixed IDs such as anthropic/claude-sonnet-4-5 across formats. Messages also accepts bare Anthropic IDs; Gemini-native accepts bare Google IDs. Those conveniences do not add models outside the catalog.
The OpenAI SDK appends the endpoint to a base URL ending in /v1. The Anthropic SDK appends /v1/messages itself, so its base URL must end at the hostname. See SDK configuration.

​
Other inputs and outputs

Use output_modality, input_modalities, and endpoints in model discovery. Sending an image-output model to Chat Completions does not generate an image; use its dedicated endpoint.

​
Conversation state and tools

Keep conversation history in your application and resend the relevant turns with each request. The gateway’s Responses endpoint does not implement stored-response retrieval or previous_response_id chaining. Do not rely on store to persist a conversation. Request logs are observability records, not conversation state. Function tools are relayed to your application for execution. Tool schemas and result formats differ between Chat Completions, Responses, and Messages. A tool result may carry an image on all three surfaces, as an inline data: base64 URI — see images in tool results. Responses also supports provider-executed search for supported xAI models; see provider-executed search.

​
Parameters and portability

  • Reasoning controls translate to the selected model’s supported controls. Their effect varies by model.
  • Prompt caching depends on the provider and prefix. A cacheable request is not a guaranteed hit.
  • service_tier accepts auto, default, fast for Claude Opus 5.5, and ultrafast for supported OpenAI models; see speed tiers. Messages uses speed: "fast" instead ("standard" is a no-op); its service_tier does not select Fast. Other tiers such as priority and flex return 400.
  • Data residency is not supported. inferenceGeo and inference_geo, including in any providerOptions bucket, return 400 on all three generation endpoints at every speed tier. Omit these controls, including on existing standard requests.
  • Models may reject sampling settings such as temperature or unsupported tool and input types. Test the exact model and request you intend to ship.
  • Do not assume OpenAI response_format or Responses text.format provides structured-output enforcement here; those controls are not part of the documented gateway contract. Validate generated data in your application.
  • Gateway-specific routing parameters from other services are not portable. Follow routing and failover for this gateway’s behavior.

​
Endpoints outside this API

Assistants, batches, fine-tuning, Files API storage, and realtime sessions are not implemented. Video input means understanding a video; it does not provide video generation. The API Reference tab contains the supported request schemas. If an SDK helper depends on a provider-specific endpoint or field outside that reference, test it explicitly before migrating it.