OpenAI introduced GPT-Live-1 and a smaller GPT-Live-1 mini, a new generation of voice models built on a full-duplex architecture: the model listens and speaks at the same time rather than alternating between a listening phase and a speaking phase. Nearly every voice assistant since Siri has worked in half-duplex, transcribing what you said, thinking, then replying, which is why interrupting them or having a natural back-and-forth rhythm has never worked well. A full-duplex model processes incoming audio continuously while generating its own speech, so it can be interrupted mid-sentence, backchannel with a quick acknowledgment while you are still talking, stay quiet while you are clearly still thinking, and adjust its response in real time instead of restarting the turn. GPT-Live also supports live web search mid-conversation and can delegate sub-tasks to other agents without breaking conversational flow. For developers, this changes what is buildable with a voice API: real-time translation that does not stall on turn boundaries, support bots that can be corrected mid-sentence instead of finishing a wrong answer, and accessibility tools that behave more like a human interpreter than a walkie-talkie. The obvious first commercial market is call centers and phone support, where the stilted turn-taking of existing voice AI is the main thing that gives it away as a bot. It shipped the same week as OpenAI's GPT-5.6 family, positioning voice -- an interface where OpenAI's consumer reach is strongest -- as a new front in its competition with Google, whose Gemini is built into Android.