OpenAI has introduced WebSocket support for its Responses API, a change aimed at reducing latency in multi-step agentic workflows [source]. The update allows developers to maintain persistent connections rather than opening new HTTP requests for each model interaction.

According to the announcement, WebSocket connections can cut round-trip overhead by up to 50% in scenarios where agents make sequential tool calls or require multiple reasoning steps [source]. OpenAI states this is particularly relevant for workflows involving code execution, web searches, or database queries where the model must wait for external results before proceeding.

The implementation uses the same authentication and rate-limiting structure as the existing REST API. Developers can upgrade an existing HTTP connection to WebSocket using standard protocol negotiation [source]. Messages are exchanged in JSON format, with the server streaming response chunks as they are generated.

OpenAI notes that WebSocket connections remain open for up to 10 minutes of inactivity before automatic closure [source]. The company recommends WebSockets for applications where agents perform more than three sequential actions, but suggests standard HTTP requests remain more appropriate for single-query use cases.

The feature is available immediately for all API tiers. OpenAI has published code samples in Python and TypeScript demonstrating connection setup and message handling [source]. The announcement does not specify whether existing SDKs will incorporate WebSocket support or whether developers must implement the protocol directly.

This marks OpenAI's first protocol-level optimization specifically targeting agentic architectures, following earlier updates focused on function calling and structured output formats.