Generate a stateless OpenAI-compatible Response through a model that advertises /v1/responses in supported_endpoints. JSON request bodies may be up to 32 MiB, including inline images.

Built-in models use their configured Responses URL. Community text models and endpoint agents registered with the Responses API use their selected URL for both Responses and adapted Chat requests. Managed prompt agents serialize Responses JSON and SSE around their configured prompt and MCP tool loop. Every other text model, including community models registered with Chat Completions, runs through the Chat Completions pipeline with the same billing, key permissions, safety, rate limits, caching and fallback; max_tool_calls, automatic truncation, top_logprobs and hosted tools return 400 on these models.

OpenAI prompt_cache_options and prompt_cache_breakpoint controls pass through direct Responses requests and Chat requests adapted to Responses. Managed prompt agents preserve caller breakpoints or apply an explicit breakpoint after their configured static prompt.

Response storage, previous response IDs, conversations, background execution, and item references are not supported; encrypted reasoning items may be replayed in input. Prompt agents execute their configured MCP tools on the server and return caller-supplied function calls for the client to execute. Replay output items with matching function_call_output results to continue. Completed MCP pairs are history, not client tool requests. Caller tool names must not start with mcp__; tool_choice supports only auto.

Successful text JSON responses and terminal streaming events contain usage; missing text-provider usage fails the response.

Metadata is passed unchanged to endpoint agents; see the agent’s documentation for supported keys.

Media models that advertise this endpoint in /models also accept text prompts. Only the last user message's text is used (or a string input on Responses); history, instructions and text-generation settings are ignored. Empty prompts, malformed Unicode and prompts consisting only of . or .. return HTTP 400. Attachments are not supported; reference-required models return their normal missing-input error. Dialogue models expect one <voice>: <text> turn per line. Images return a Markdown image embed; audio, video and 3D return a Markdown link. Both include the plain public URL. With stream: true, events are emitted after generation finishes. Media is billed normally, without text-token usage: Responses returns usage: null; Chat JSON omits usage, while Chat streaming chunks contain usage: null and have no final usage chunk. Use the native media endpoints for generation settings.