More natural voice
OpenAI launched GPT-Live-1 and its mini version, a new generation of voice models for ChatGPT. The interaction is now more like a human conversation: the system listens, responds and even pauses when the user thinks.
During the live performance, the voice mode responded immediately, maintained rhythm, and displayed natural intonation. He also performed simultaneous translation, listening and speaking at the same time.
According to OpenAI, more than 150 million people use their voice weekly on ChatGPT to answer questions, practice languages or drive.
Full-duplex architecture
The main novelty is the full-duplex architecture: the model can listen and speak at the same time. It no longer waits for the user to finish before responding, but instead intervenes when appropriate or remains silent if it detects that the person is still thinking.
“GPT Live sounds so natural that it’s almost like talking to a real person. It’s a very special experience that you have to try to feel,” said Atty Eleti of the OpenAI technical team.
The system avoids interruptions due to environmental noise or short pauses. You can now use expressions such as “uh-huh” to indicate that you are still attentive. The user can ask to speak more slowly, pause or interrupt with new questions without breaking the flow.
Simultaneous translation and reasoning
When a query requires internet searching or complex analysis, GPT-Live delegates that task to GPT-5.5. The conversation continues without interruption while the frontier model processes the response.
“When GPT Live receives a question that requires deep analysis, it can delegate that reasoning to GPT-5.5, while GPT Live continues conversing with the user,” explained Kundan Kumar of the technical team.
Users choose between three levels of reasoning: Instant, Medium and High. Answers may include visual cards when it is more helpful to show than to tell.
Simultaneous translation works thanks to the full-duplex architecture: you listen to one language while generating the translation in another, similar to a professional interpreter.
Security and availability
OpenAI reinforced security with specific training for voice conversations and mechanisms that act in real time. If it detects risky content, redirect the conversation or end the interaction. The system only uses predefined voices, without imitating real people.
The update will be available this week for iOS, Android and ChatGPT.com. GPT-Live-1 will be the default model on Go, Plus and Pro plans; GPT-Live-1 mini is coming to the free plan.




