OpenAI has released GPT-Live-1, a new voice system for ChatGPT that lets the app listen and talk at the same time, instead of waiting for you to finish speaking before it responds. The update is available now on iPhones, Android phones, and computers The Verge.

The old ChatGPT voice mode worked more like a game of tennis — you hit the ball, then waited for your opponent to hit it back. GPT-Live-1 is different. It can process your words while it's already speaking, much like a real conversation between two people. This means you can interrupt it mid-sentence, and it can respond naturally instead of freezing or getting confused.

How it works

OpenAI calls this "full duplex" processing, which is engineering jargon for two-way communication that happens at the same time. One of OpenAI's product leaders explained it this way: the system can speak and listen simultaneously, rather than taking turns The Verge.

The model also adds human touches like saying "yeah," "got it," or "mhmm" while you're talking — small verbal nods that signal it's paying attention without actually interrupting you. If you ask ChatGPT Voice to stop talking and listen, it can do that now. The old version couldn't.

Real-time translation is part of this release too. GPT-Live-1 can start translating what you're saying into another language while you're still speaking, rather than waiting for you to finish your sentence first. This puts it in more direct competition with specialized translation apps, though OpenAI hasn't released detailed speed or accuracy comparisons yet.

Who gets it and what it costs

This new voice model powers ChatGPT for paying subscribers — Go, Plus, and Pro tiers. People using the free version of ChatGPT get GPT-Live-1 mini, a smaller and less powerful variant that uses less computing power. OpenAI's decision to tier access suggests that handling voice and interruptions at the same time is more expensive and demanding than generating text responses, which lines up with how other companies in the industry experience voice AI — it requires more computing power per conversation than text does.

Safety built in

OpenAI has added safeguards to the voice model to prevent harmful responses. If the conversation involves self-harm, the system is designed to provide crisis helpline information. It also gives age-appropriate answers when talking to teenagers. The company can end conversations entirely in what it calls higher-risk situations The Verge.

Voice assistants carry different risks than text chat. When an AI listens continuously, sounds natural, and feels like it's actually present with you, people treat it differently. They may talk to it in moments they wouldn't type to a computer — when they're lonely, upset, or in crisis. OpenAI's choice to build in detection for self-harm and teen-specific responses suggests the company expects people to use this voice mode the way they'd use a trusted listener, not just as a tool.

The harder engineering problem OpenAI solved here is making a model listen and speak at the same time. That's not easy — the system has to schedule incoming and outgoing audio to overlap, rather than just trading turns like most voice assistants do. OpenAI chose not to build all the reasoning power directly into the voice model. Instead, when you ask something complex or ask it to search the web, the voice layer hands off the request to GPT-5.5, a more powerful text model, then reads the answer back to you.

That choice tells us something about where the technology stands. Building a voice assistant that can interrupt and be interrupted is one hard problem. Building one that reasons deeply about complex questions is another. For now, OpenAI is solving them separately rather than trying to do everything in one model.