AI voice modes have been so frustrating that most people don’t bother with them. The voice modes for ChatGPT, Claude, and Gemini have relied on turn-taking — one side listens while the other talks — creating awkward, jerky conversations filled with long pauses and abrupt interruptions.
OpenAI is now rolling out an upgraded chat mode called GPT-Live that aims to fix this. The key change is a full-duplex architecture, which allows the model to both listen and talk at the same time. It’s rolling out now to all ChatGPT users, including free users.
Previous ChatGPT voice modes worked like old CB radios. When you talked, the AI listened. When the AI talked, you waited. GPT-Live changes that by working more like a standard phone call — you can interrupt, talk over the AI, or just let it keep going.
Because the system “continuously processes input while generating output,” it’s better at deciding when to talk, when to interrupt, and when to just stay quiet. That last part — knowing when to shut up and listen — is something earlier voice modes struggled with.
GPT-Live can also deploy agents to conduct web searches while it’s talking. This eliminates the awkward “searching the web” pauses that plagued the old voice mode when the AI needed to look something up.
OpenAI isn’t the first to offer this. AI voice startup Thinking Machines has demonstrated full-duplex voice interaction models. There’s also Sesame AI, which has a voice mode capable of deploying search agents while still speaking.
But for most users, GPT-Live will be the first real encounter with an AI that can hold a fluid, two-way conversation. They are releasing two versions: GPT-Live-1 and GPT-Live-1 mini. The writer of the report hasn’t tried it yet and will share first impressions once the update lands on their account.
Voice remains the most natural interface for many tasks — driving, cooking, walking — but only if the system feels like a real conversation partner rather than a walkie-talkie. The full-duplex approach brings AI voice closer to how humans actually talk, where interruptions, overlapping speech, and simultaneous listening are normal, and it can even discover behavior in a more human-like way.
Turn-based voice modes forced users into a rigid pattern: speak, wait for response, speak again. That works for simple commands but fails for anything resembling real conversation. You couldn’t interject a clarification, correct a misunderstanding mid-sentence, or ask a follow-up without waiting for the AI to finish its monologue.
The full-duplex architecture changes the dynamic entirely. It can hear you while it’s still forming its own response. It can adjust mid-sentence based on what you just said. It can also decide to let you finish your thought without jumping in — a skill that sounds trivial but is surprisingly hard for AI models to learn.
The web search agent feature is a practical addition. Previously, if ChatGPT needed to look something up, the conversation halted entirely. Now it can search and talk at the same time, keeping the flow going, much like how mobile professionals need to stay connected and informed on the go.
GPT-Live is rolling out starting today.
The update hasn’t appeared on the writer’s account yet.