Long reasoning on streams
- Streaming requests open immediately and send keep-alive events while the model reasons, so long reasoning no longer ends with HTTP 524.
- ChatGPT steps no longer have a gateway time limit by default.
Changelog
Product updates and API improvements.
Stored responses for ChatGPT connections
- store: true stores a response, encrypted at rest, for 30 days; requests without it are not stored.
- previous_response_id continues a stored response with only the new input, keeping the model reasoning.
- GET /v1/responses/{id}, GET /v1/responses/{id}/input_items and DELETE /v1/responses/{id} are available.
- The privacy policy describes stored responses and conversation context retention.
Codex CLI and Responses compatibility
- Codex CLI can use ChatGPT connections: additional_tools items, tool namespaces and custom tools such as exec are supported.
- Valid OpenAI parameters without a ChatGPT-account control, such as client_metadata, are accepted and reported in x-gateway-ignored-parameters instead of returning 400.
- text.verbosity is applied, and inference requests may be up to 16 MiB.
- Documented how code mode (tool_mode: code_mode_only) changes tool calls.
Model availability per connected account
- A model that a connected account no longer offers returns 400 model_not_available instead of a generic gateway error.
- When another of your connected accounts offers that model, the request is served there without a retry from your client.
- Only the account that was rejected stops listing the model; your other accounts and other users are unaffected.
Responses API parity for ChatGPT connections
- Encrypted reasoning items in Responses, kept across tool calls and when a conversation context is rebuilt.
- New POST /v1/responses/compact endpoint and automatic compaction through context_management.
- Token usage on every response, including tool-call steps and streams, with cached and reasoning token details.
- Tool calls issued together are returned together; parallel_tool_calls=false delivers them one per step.
- Retry-After on every usage limit response and a dedicated 504 gateway_timeout code.
- A longer time limit for long reasoning steps and failed Responses streams reported as response.failed.
- Documented agent mode and that model limits come from the OpenAI catalog of each connected account.
Stored responses and previous_response_id remain unavailable for ChatGPT connections. See ChatGPT Responses parity.
API integration improvements
- Improved continuity for supported tool workflows.
- More consistent availability responses and retry guidance.
- Optional bounded waiting for busy requests.
- Improved compatibility with caching options and model controls.
- Clearer model capabilities, usage reporting and request tracking.
- Additional improvements for agent integrations.
- Updated reasoning controls and conversation context retention.
- Improved cache reuse across independent conversations and integration compatibility.
Available options depend on the selected connection and model. See the API documentation for the current contract.