OpenAI announced that it has made its Codex agent loop 40% faster by introducing WebSocket support and connection‑scoped caching to the Responses API. The change allows the system to keep a persistent connection and reuse conversation state, tokenization, and safety checks across multiple turns, eliminating the need to rebuild the full history for each request. During a two‑month sprint, the team built a WebSocket mode that kept an in‑memory cache of previous response state and reused rendered tokens, safety classifiers, and routing logic. Alpha testing with key coding‑agent startups reported up to 40% improvements in agentic workflows. After launch, Codex shifted most of its traffic to WebSocket mode, achieving the 1,000 tokens‑per‑second target for GPT‑5.3‑Codex‑Spark and bursts up to 4,000 TPS. The performance gains were also seen by external partners: Vercel’s AI SDK saw up to 40% latency reduction, Cline’s multi‑file workflows improved by 39%, and OpenAI models in Cursor became up to 30% faster. These results demonstrate that reducing API overhead through persistent connections is essential as inference speeds increase, enabling smoother, lower‑latency agentic experiences for developers.
Speeding up agentic workflows with WebSockets in the Responses API
OpenAI accelerated its Codex agent loop by 40% using WebSockets and connection‑scoped caching, enabling the new GPT‑5.3‑Codex‑Spark model to reach 1,000 tokens per second. The shift from synchronous HTTP calls to a persistent WebSocket connection reduced API overhead and allowed reuse of conversation state, tokenization, and safety checks. Alpha testing with startups and subsequent launch saw latency drops of up to 40% for developers and 39% for Cline’s multi‑file workflows.
Why it matters
Reducing API overhead is critical as LLM inference speeds up; persistent connections enable real‑time agentic workflows and lower operational costs for developers.
Topic dossiers OpenAI