API specification

Send a first request, select a model from the catalogue, and retain a request identifier for diagnosis.

Send an API key in a Bearer header and keep production credentials separate from development credentials.

Use different base URLs for OpenAI-style and Messages clients: they target one specification but compose paths differently.

Configure a client with a key, model, and base URL that matches its protocol; keep the secret in the environment.

Short practical answers about keys, base URLs, models, cost, and integration diagnosis.

Build model selection from catalogue data instead of a hard-coded list in your application.

Separate keys, owners, budgets, and usage traces across applications and workspaces.

Treat key, organization, model, and routing limits as part of the user workflow.

Connect response metadata and your own request logs to real user actions.

Minimize transmitted data, separate an end-user identifier from content, and control retention in your product.

Create multi-turn conversations in the familiar messages format and select a model for each request.

Use the Messages format for conversations with a top-level system instruction and an explicit output limit.

Use one input object when your application benefits from handling output items and actions in a single data model.

Receive tokens and structured deltas through SSE without waiting for the complete response.

Describe functions with JSON Schema, execute the call in your application, and return the result to the conversation.

Request a JSON object or JSON Schema when software, rather than a person, will consume the response.

Set reasoning effort per request only for models that declare the capability.

Express execution requirements in provider preferences and retain the actual outcome for observability.

Separate recurring context so a client can use compatible caching where the chosen model supports it.

Retrieve fresh sources through a client-executed `web_search` function and keep links separate from generated text.

Upload a file through a dedicated operation, retain its id, and pass a reference in a later compatible request.

Create images through one route and select parameters from the capabilities of the chosen model.

Start video as an asynchronous job, retain its job id, and keep creation separate from reading the result.

Use separate operations for speech synthesis, transcription, and audio input to a model.

Transform text or supported inputs into vectors for search, clustering, and recommendations.

Reorder a small set of candidates against a query to improve the precision of final retrieval.

Process independent operations in controlled groups when the workflow permits asynchronous delivery.

Connect an AI client to permitted tools, resources, and prompts through a connection URL issued in account settings.

Build handling around HTTP status, a structured error body, and request id rather than message text.