The OpenAI Responses endpoint

Alongside Chat Completions, the gateway also serves OpenAI's Responses API:

This is the endpoint the Codex CLI and the OpenAI SDK's client.responses.create() use.

Authentication and address: the same as Chat Completions

Authorization: Bearer your key

The address is the same one too (with /v1):

https://api.xisu.ai/v1

Key, tenant, balance, plan and budget are shared with the other two surfaces — nothing extra to request, and usage lands on the same account.

Request and response

Standard Responses shape, with model and input required. If you send max_output_tokens, it must be a positive integer. Apart from the model name, the request body reaches the upstream as-is — reasoning, instructions, function tools and the rest are not translated.

Responses are returned as-is. Streaming is a native Responses event stream (response.created … response.output_text.delta … response.completed) with no data: [DONE], so don't write your termination check against it. Usage appears once, in response.completed, as input_tokens / output_tokens (not prompt_tokens / completion_tokens).

⚠️ Three behaviours that differ from the official API

1. No stored conversations: three fields are rejected with 400

The gateway does not store your conversations. These fields depend on server-side storage and return 400 instead of being silently ignored:

(Ignoring them silently would send the upstream a request with no history — the model answers out of context and you are still billed. So we fail loudly instead.)

2. Server-side tools are billed per call

Tools executed by the upstream — web_search, file_search, code_interpreter, mcp and the like — pass through as-is but are billed per call, outside token usage.

• They are allowed only when the model has a per-call price configured; otherwise, or if we don't recognise the tool type yet, you get 400 server_tool_not_priced;
• Tools you execute yourself, such as function and custom, are not billed per call and pass through normally;
• Requests that use per-call tools are not covered by prepaid packs; they are charged to your balance in full.

3. The models here are a subset

Only models with an OpenAI-compatible upstream can be called on this endpoint. The two failure cases call for opposite next steps:

Models marked OpenAI Responses in the model catalog and the console's available-models list can use this endpoint; an unmarked model is not necessarily unsupported — the actual call is the source of truth. All other error codes are the same as for Chat Completions; see Error codes and rate limits.