What's this about?
OpenAI rebuilt the entire voice pipeline behind GPT-Live, ChatGPT's voice mode, so the model can keep audio flowing continuously — it can hear you interrupt while it's mid-sentence, and it can run slow reasoning or tool calls in the background without the conversation going dead.
What exactly is "full-duplex"?
Borrowed from telephony: half-duplex is a walkie-talkie (one side talks, the other waits), full-duplex is a phone call (both sides can talk and hear at once). Voice assistants until now were walkie-talkies with extra steps — record you, stop, think, then play a reply. A full-duplex model streams both directions simultaneously, which is what makes barge-ins ("no, not that one—") feel natural.
What's the community saying?
5.5k likes on the announcement; the HN story is still young (posted late in the day).
- The engineering angle: keeping audio alive while deep reasoning and tool use run means voice no longer has to choose between "fast but shallow" and "smart but awkwardly silent".
- Skeptics note "listen while speaking" demos well but the test is messy real-world audio — crosstalk, background TV, two people talking.
- Builders are asking when the rebuilt stack lands in the Realtime API rather than just the ChatGPT app.