GPT-Live shifts ChatGPT Voice from turns to continuous audio

Voice AI moves onto a continuous path

OpenAI’s GPT-Live post is less about a new voice demo and more about the systems work required to make conversation feel live at ChatGPT scale. The company wrote on X that GPT-Live can listen while it speaks and that OpenAI rebuilt the voice stack from the client to the model so deeper reasoning and tool use do not interrupt the conversation.

GPT-Live can listen while it speaks. To make that feel natural at ChatGPT scale, we rebuilt the voice stack from client to model.

The linked engineering article explains why earlier voice systems struggled. Turn-based pipelines had to decide when a user had finished speaking before sending work to the larger model. GPT-Live instead uses a full-duplex voice model, keeps audio on a dedicated fast path, and lets deeper reasoning or tool calls happen asynchronously. That separates the real-time media loop from application logic.

The concrete latency number is the startup path. OpenAI says its WebRTC Abridged Roundtrip Protocol and Instant Connect reduce media and data startup from six network round trips to one. The post also describes moving media frontend and inference logic to Go, managing stateful inference across long sessions, and preparing replacement model instances before context compaction or handoff would otherwise create a pause.

OpenAI’s account typically signals product and research milestones, but this tweet points to infrastructure that could matter for the upcoming GPT-Live API. The next thing to watch is whether developers get enough control over the fast path, delegation policy, and observability to build real-time voice agents that remain responsive outside OpenAI’s own clients. Source tweet