Sign in
inference

`/v1/responses` — OpenAI Responses API (MVP, stateless single-turn).

OpenAI Responses API (MVP). Strict allow-list parser (product/design.md §5.14): unknown top-level fields reject with 400 unsupported_field. input is either a string prompt OR an array of {role, content} messages (Workstream A array-input support). Known-rejected fields: tools, tool_choice, previous_response_id, reasoning, background, metadata, instructions — each rejects with 400 unsupported_field. stream: true is rejected (Responses SSE is deferred). Multimodal image content parts reject with 400 unsupported_field on the array form.

post/v1/responses

Request body

max_output_tokensinteger nullable

Defaults to 16 (mirroring completions) when absent. Positive integer.

modelstring required
seedinteger nullable

Best-effort determinism seed (i64 reinterpreted as u64). Non-integer values reject with 400 invalid_request.

streamfalse | null nullable

Accepted only as false (or absent). true rejects with 400 unsupported_field (Responses SSE is deferred; tracked in product/design.md §5.14).

temperaturenumber float nullable

Sampling temperature. Finite number >= 0.

top_pnumber float nullable

Nucleus sampling. Finite number in (0, 1].

Response

Response object

Changes