OpenAI launches GPT-Live: full-duplex voice AI that listens and speaks simultaneously
OpenAI launches GPT-Live: an AI that listens and speaks simultaneously like a human
OpenAI has just made a giant leap in voice interaction with artificial intelligence. GPT-Live, announced today, is the first generation of voice models with a full-duplex architecture — capable of listening and speaking simultaneously, without that awkward pause while the model processes.
If you've used ChatGPT Voice before, you know the drill: you speak, stop, wait for the spinning dot, and then the response comes. GPT-Live breaks this cycle. It can interrupt, say "hmm", "yeah", "I see" while you're still talking — just like a real conversation.
How it works in practice
The full-duplex architecture means the model processes input and output audio in real-time, without relying on a cascade of models: transcription → LLM → text-to-speech. Instead, GPT-Live has a unified pipeline that handles conversation continuously.
For tasks requiring more firepower — web search, deep reasoning, complex processing — the model delegates to GPT-5.5 in the background. Meanwhile, it keeps talking naturally with you, and when the response is ready, it's inserted into the conversation seamlessly.
What changes for users
The experience is dramatically more natural. Demos show the model performing live simultaneous translation, participating in brainstorming sessions where it suggests ideas mid-sentence, and even handling calls where it asks you to hold on while it processes something.
Two versions are launching: GPT-Live-1 (full) and GPT-Live-1 mini (lighter). Both are available globally in ChatGPT, with API access coming soon.
Want to try GPT-Live? It's already running in ChatGPT. API version has a waitlist open.