OpenAI's WebSocket API mode speeds agents 40%

OpenAI's performance sprint on the Responses API, starting around November 2025, initially cut time to first token by close to 45% through caching and fewer network hops. But that wasn't enough for the new GPT-5.3-Codex-Spark model, which targets over 1,000 tokens per second—an order of magnitude faster than prior flagship models like GPT-5. The deeper problem was structural: each request was treated independently, forcing repeated processing of the entire conversation history even when little had changed.

The solution meant ditching the pattern of many HTTP calls for a single persistent WebSocket connection. An internal prototype showed the potential, and after a two-month build, OpenAI launched a version that keeps the developer-friendly API shape with `response.create` and `previous_response_id`. The server caches prior state in memory, so only new input gets validated, classified for safety, and tokenized, avoiding redundant work across multi-turn agent rollouts.

The public launch has seen immediate adoption, with Codex moving most of its traffic to the new mode and achieving up to 4,000 TPS bursts in production. External partners like Vercel, Cline, and Cursor also report up to 40% latency improvements. The key question is whether broad adoption follows, since the benefit hinges on developers choosing to integrate WebSocket mode to see these gains.