Speeding up agentic workflows with WebSockets in the Responses API (openai.com)

OpenAI has announced a significant upgrade to its Responses API with the integration of WebSockets, enhancing the speed and efficiency of agentic workflows for its coding models like GPT-5.3-Codex-Spark. This update allows for a dramatic increase in latency performance, boosting token generation speeds from 65 tokens per second to over 1,000 TPS, and even achieving bursts of up to 4,000 TPS. By enabling a persistent connection and caching state in memory, the API reduces repetitive processing and unnecessary network calls, leading to a more responsive experience for users, particularly when executing tasks like bug fixes and code suggestions. This development is critical for the AI/ML community as it addresses the growing demand for faster inference driven by increasingly optimized models. The use of WebSockets not only minimizes API overhead but also simplifies integration for developers, making it easier to adopt this new communication protocol without extensive modifications to existing workflows. Early adopters have reported latency reductions of up to 40%, indicating that this enhancement could significantly improve productivity and real-time responsiveness in AI applications. Overall, this upgrade represents a pivotal step in ensuring that surrounding services keep pace with rapid advancements in model capabilities.