Agent inference

OpenAI Responses API

What it is

Source: OpenAI, Responses API reference, read 2026-09-25

Source: OpenAI, Codex configuration reference, read 2026-09-25

Source: RunInfra, OpenAI compatibility, read 2026-09-25

Source: RunInfra, Codex configuration, read 2026-09-25

Why it moves cost and latency

Codex custom providers use model_providers.<id> with wire_api = "responses", the only supported wire value and the default when omitted. RunInfra's endpoint is a stateless Chat Completions compatibility adapter, so Responses support does not imply stored conversations or hosted tools.

What it looks like in practice

RunInfra's Codex configuration uses [model_providers.runinfra], base_url = "https://api.runinfra.ai/v1", wire_api = "responses", and env_key = "RUNINFRA_API_KEY". POST /v1/responses accepts flat or nested function-tool declarations, streams the item lifecycle from response.output_item.added through response.completed, and returns reasoning as a reasoning output item. Clients resend the conversation because previous_response_id and truncation: "auto" are unsupported; store is treated as false, and conversation storage, hosted tools, and background jobs are not implemented.

Related terms

Questions this definition answers

What does OpenAI Responses API mean in inference serving?

The OpenAI Responses API is an interface for generating model output from inputs and representing text, reasoning, and tool calls as response items. Its request and event shapes differ from Chat Completions, so a client must use the protocol its provider implements. RunInfra's Codex configuration uses [model_providers.runinfra], base_url = "https://api.runinfra.ai/v1", wire_api = "responses", and env_key = "RUNINFRA_API_KEY". POST /v1/responses accepts flat or nested function-tool declarations, streams the item lifecycle from response.output_item.added through response.completed, and returns reasoning as a reasoning output item. Clients resend the conversation because previous_response_id and truncation: "auto" are unsupported; store is treated as false, and conversation storage, hosted tools, and background jobs are not implemented.