| Method and path | POST /responses |
| Purpose | The same router in the OpenAI Responses dialect, the default transport for the OpenAI Agents SDK and Pydantic AI. |
| Authentication | API key or session |
| Limits | As chat completions. |
Request
input takes the string form or the item-list form, including a full function_call and function_call_output transcript. instructions, tools, tool_choice, text.format, and reasoning.effort map onto the routed request. model takes the same values as chat completions, auto included.
The endpoint keeps no conversation state: a response cannot be retrieved or continued by its id, so each call carries the whole conversation. This is separate from the request log, which keeps request content for up to 30 days unless it is switched off; see What Inferbase stores from your requests.
A request names a session with the X-Inferbase-Session header and an image answer with the X-Inferbase-Intent: image header, since the Responses shape has no routing field. The image arrives as an image_generation_call output item whose result is the base64 image, as OpenAI returns it. When the provider answers with a URL, Inferbase fetches the image and inlines it, and a failed fetch is a 502 saying the image was generated but could not be delivered. An image answer is non-streamed only. The speech and video intents are refused with a 400, because this dialect has no shape for them. Without the header the answer is text and no media intent is inferred.
| Field | Behavior |
|---|---|
store | Accepted and ignored: no response object is kept for later retrieval. The response reads back store: false. The request log is governed by the key and zero retention settings, not by this field. |
previous_response_id | Refused with a 400. A request that believes ten turns are in context when one was sent is a corrupted loop, and failing loudly is the smaller harm. Send the whole conversation in input, which the Agents SDK does by default. |
Hosted tools: web_search, file_search, code_interpreter, mcp | Refused by name. Inferbase routes and serves models; it does not run OpenAI's hosted tools. |
Example
curl https://api.inferbase.ai/api/v1/inference/responses \
-H "Authorization: Bearer inf_your_api_key" \
-H "Content-Type: application/json" \
-d '{
"model": "auto",
"instructions": "You are a concise assistant.",
"input": "List three uses of a rerank model."
}'Response
The Responses shape. Streaming emits the full typed event sequence. The served model and the request id carry the same meaning as on chat completions.
Errors
As chat completions, plus a 400 for previous_response_id or a hosted tool type. See Errors.
Last updated October 10, 2026.