OpenAI rebuilt the voice stack so GPT-Live can listen while it speaks

On August 3, 2026 OpenAI published an engineering post on how it built a realtime system for GPT-Live, ChatGPT's third-generation voice stack, in six months. The core change: the voice model is full-duplex, can listen and speak at once, and the separate turn detector is gone from the audio path. Deeper reasoning and tool use go asynchronously to a frontier model (GPT-5.5 in the examples) without having to interrupt speech.

A full-duplex model runs the call; the turn detector no longer sits in the way

Older voice stacks inherited turn-based logic from text LLMs: a detector first decided whether the user had finished, then the large model could start. A bad guess either cut the user off or added lag. GPT-Live puts conversation control on the voice model: audio streams in and out continuously, while delegation, tool use, and persistence run off the live path.

To hold that at ChatGPT scale, OpenAI split the media frontend from application logic. Audio gets a dedicated fast path; a slow tool call must not stall frame time. The media frontend moved from Python asyncio to Go; the new path's p95 smoothness reportedly matches the old p50. Transport stays on WebRTC, with WARP (WebRTC Abridged Roundtrip Protocol) and Instant Connect added to cut network round trips before the first audio frame.

For product teams, voice becomes an agent channel, not just TTS chat

Help Center confirms the product fallout: Live runs on GPT-Live-1 for paid plans and GPT-Live-1 mini on free. It supports full-duplex, web search, memory, plus text and images in the same chat. At launch it does not support video, screen sharing, connected apps, or plugins; those remain on Advanced voice where they still matter.

More important than a "nicer conversation" is the architecture for Voice in Work and Codex in the desktop app: start tasks by voice, check progress, and coordinate multiple agents in one call. Voice stops being only an accessibility feature and becomes a control surface. Limits are real and plan-shaped: Pro gets much higher quotas, Go/Plus roughly hours, free only limited Live mini, and Business/Enterprise often burn credits (for example 5 credits per minute on flexible pricing). One Live conversation caps at 2 hours.

A full-duplex demo is not yet a reliable meeting agent

OpenAI itself admits segmentation limits: the UI still needs discrete turns while the live path is continuous. The server therefore keeps speculative and authoritative views of who holds the floor, so a short "mm hmm" does not junk history. Overlap, noise, and multi-speaker rooms still break attribution. Live is primarily one-on-one, not a meeting recorder for five people at a table.

The second ceiling is product scope. Without video/screen share and without plugins at launch, Live does not cover scenarios Advanced still owns. Enterprise/Edu also need Voice and Early Model Access enabled during the early-access window. Anyone expecting "everywhere immediately" will hit admin policy and quotas before they hit the model.

Interruption quality and delegation under load will show if it is truly live

Three signals matter. First, whether full-duplex keeps interruption quality outside lab demos (coughs, overlap, weak mics). Second, whether async delegation to GPT-5.5 stays inside a human-acceptable latency budget when the tool call is not trivial search. Third, Voice in Work/Codex adoption: whether desktop users actually steer agents by voice, or fall back to text the moment the brief must be auditable.

If those three hold, OpenAI is not selling smoother TTS. It is selling voice as a primary runtime for agent work. If they do not, this stays a nicer phone call with a chatbot, and Advanced keeps the specialized inputs.

Lilith's verdict

OpenAI did not win this with a prettier voice. It won by splitting talking from thinking: audio stays on the fast path while search and tool use finish on the side without breaking the call.