OpenAI has published the engineering design behind GPT-Live, its third-generation voice system for more natural, responsive conversations. The system replaces the familiar stop-and-start pattern of voice assistants with a full-duplex model that can listen while it speaks, then delegate harder work to a frontier model without pausing the audio exchange.
The architecture already underpins ChatGPT Voice and, according to OpenAI, supports the newly launched ability to control a computer and coordinate agents in the ChatGPT desktop app. It is also intended to support a forthcoming GPT-Live API, although OpenAI has not announced an API release date, price or detailed availability.
A conversation loop without a turn detector
Earlier voice systems typically waited for a small turn-detection model to decide that a person had finished speaking. If it guessed too early, the assistant interrupted; if it waited too long, the exchange felt slow. Only after that decision could the larger language model begin preparing a response.
GPT-Live removes that detector from the audio path. Incoming speech streams continuously to the voice model and generated speech streams back at the same time. OpenAI says the result is sub-second responsiveness and a model that can use tone, pacing and overlap as part of the conversation rather than reducing each turn to a separate audio file.
That is an architectural shift rather than a single latency tweak. The primary system is responsible for keeping the media loop moving. Reasoning, tools, persistence and other application work sit behind an asynchronous boundary, so a slow database lookup or tool call should delay its own result without freezing the voice.
Two models can contribute to one exchange
The fast voice model is not expected to do every task alone. When a request needs search, deeper reasoning or a tool, GPT-Live can consult another model, including GPT-5.5, while continuing the spoken interaction. OpenAI describes this as separating the job of talking from the job of thinking.
To keep delegation useful in a live conversation, the second model and its tools are prepared when a session begins. Its initial context is prefilled, later requests use session affinity and prompt caching, and settings such as reasoning effort and output length are managed as part of the response-time budget. The voice model can briefly carry the exchange, but it cannot disguise an indefinitely slow tool.
Continuous speech still has to be represented as messages for ChatGPT’s interface, analytics and safety systems. The application server therefore maintains a provisional view of the current exchange while building an authoritative transcript for systems that need stable records. OpenAI says this lets the display remain current without forcing the model back into rigid turns.
State moves without stopping the call
Long voice sessions create a different operational problem: the context grows while model instances need to scale up, scale down or be replaced. OpenAI’s hand-off mechanism warms a replacement instance, fills it with the current conversation state, briefly runs old and new instances in parallel, and switches only when the replacement is ready.
The same approach is used for context compaction. Rather than pause while an overlong conversation is condensed and its attention cache rebuilt, the active instance keeps speaking while a replacement is prepared with the smaller context. That design aims to make infrastructure changes inaudible to the person on the call.
Latency work reaches the network handshake
OpenAI rewrote its media front end and inference logic in Go, replacing a Python asyncio implementation. The company reports that the new system’s 95th-percentile frame-delivery performance matches the median of the previous implementation. WebRTC remains the transport layer because it can cope with packet loss, clock drift and reconnects.
The team also developed WebRTC Abridged Roundtrip Protocol, or WARP, a set of proposed improvements designed to reduce session start-up from six network round trips to one. OpenAI says support is already present in libwebrtc and Pion, with the proposals moving through an IETF working group. A separate Instant Connect mechanism pre-negotiates session parameters so a valid session can begin from the first UDP packet, while retaining the ordinary signalling path as a fallback.
Production traffic shaped the final system
Before users heard GPT-Live, OpenAI shadowed a gradually increasing share of real ChatGPT Voice sessions. The existing voice experience served users while the new path ran in read-only mode. The trial exposed bottlenecks that short load tests had missed, including CPU-side stream handling, geographic routing, long-session memory pressure and races during disconnects.
The account is useful because it identifies the limits as clearly as the advances. Responsive voice depends on the whole path, not only model inference, and the deeper model still has to return quickly enough to rejoin the conversation. OpenAI has not provided independently verified latency results, supported-device details for the forthcoming API or a deployment timetable. For now, the announcement explains the technical foundation behind an evolving product rather than a complete developer launch.