OpenAI WebSocket mode cuts agent-loop latency 40%

Hot take: this is the kind of infrastructure upgrade that only becomes obvious once the model itself stops being the slow part.

• –The real win is not WebSockets as a transport; it is state reuse and less repeated work across tool-call loops.
• –This targets a structural latency problem in agentic systems, so the payoff compounds on long, multi-step workflows.
• –It is especially relevant for coding agents, browser automation, and orchestration layers where repeated round trips dominate perceived speed.
• –The launch is also a signal that OpenAI expects fast inference to keep forcing the surrounding API stack to get leaner.