ChatGPT's New Voice Mode Can Listen and Talk at the Same Time
OpenAI's GPT-Live uses a full-duplex architecture to make voice conversations feel human — and quietly hands hard questions to GPT-5.5 in the background.
OpenAI has released GPT-Live, a new generation of voice models for ChatGPT that can listen and speak at the same time — the way people actually talk. Until now, voice AI worked like a walkie-talkie: you speak, it waits, it answers. GPT-Live uses what engineers call a full-duplex architecture — meaning both directions of the conversation stay open simultaneously, like a normal phone call.
Two versions are rolling out worldwide. GPT-Live-1 is for paying users on the Go, Plus, and Pro plans; GPT-Live-1 mini comes with free accounts. Both work on iOS, Android, and chatgpt.com, with API access for developers planned. The model decides multiple times per second whether to speak, keep listening, pause, or interrupt. It even uses small filler sounds like “mhmm” to signal it’s following along — and you can cut it off mid-sentence without breaking anything.
The cleverest part is hiding in the background. Live voice models have historically been fast but not very bright — great at chatting, bad at facts. GPT-Live fixes this by delegating: when a question needs a web search or real reasoning, it quietly hands the task to GPT-5.5 while keeping the conversation going. The numbers are striking. On GPQA, a test of scientific reasoning, GPT-Live-1 scores 84.2 percent where the old Advanced Voice Mode managed 45.3. On BrowseComp, a web-research benchmark, the gap is 75.2 percent versus 0.7.
Here’s what’s behind it: OpenAI is splitting the job in two. One model handles the rhythm and social texture of conversation, another handles the thinking. That mirrors how the whole industry is moving — small fast models up front, big smart models behind the curtain. Nvidia released a similar open-source model called PersonaPlex earlier this year, so the approach is catching on.
A fair caveat: the more human an AI sounds, the easier it is to forget it isn’t one. Research has linked heavy voice-mode use to emotional dependency, and OpenAI itself ships GPT-Live with extra safety layers — crisis hotlines, parental controls, and the ability to end conversations in high-risk situations. Voice with video and screen sharing isn’t supported at launch, though OpenAI says it’s coming.
What this means for you: If you’ve found AI voice modes awkward — that stiff pause, the robotic turn-taking — this is the update that might change your mind. Free users get the mini version, so anyone can try it. It’s genuinely useful for hands-free moments: cooking, driving, walking through a problem out loud. Power users get a voice interface that can finally do real research mid-conversation. Just keep one thing in mind: it sounds more human now, but it’s the same kind of software underneath.
Sources
Where the Network Ends, Small AI Begins: Tiny Models Are Saving Lives Offline
IEEE Spectrum reports how phone-sized AI models authenticate medicines, spot crop disease, and run ECGs in places with no broadband and no data centers.
