OpenAI launches new voice model that works ‘like a phone call’

Image credit: OpenAI

OpenAI has launched a new voice model dubbed GPT-Live-1. The AI firm will begin rolling out the model today across ChatGPT’s web and mobile apps for both free and paid users globally.

According to the company, more than 150 million people use voice to interact with the chatbot every week, either through dictation or voice mode.

GPT-Live-1 will replace Advanced Voice Mode as the default model for ChatGPT’s voice chat feature. OpenAI described the new iteration as a full-duplex model, which means it can listen and speak at the same time.

GPT-Live-1 is OpenAI’s first voice model with that capability, the company said in a press briefing. It works “more like a phone call,” said Kundan Kumar, who leads research for speech models at OpenAI. Previous versions worked more like walkie-talkies, where one side had to finish speaking before the other could respond.

With the new iteration, OpenAI said talking with ChatGPT should feel “much more like talking with another person.” GPT-Live-1 can say things like “uh huh” or “right” to show that it’s paying attention. It can also jump in proactively to correct mistakes or stay quiet when users are pausing to gather their thoughts.

Other companies are also working on full-duplex models. The most notable example is Nvidia, which released its PersonaPlex system under open-source licenses in January.

Other examples include ByteDance’s Seeduplex and Interaction Models from Thinking Machines Lab. The latter is still in preview and hasn’t been widely released.

OpenAI said ChatGPT’s voice chat feature is now more intelligent by drawing on the firm’s latest frontier model, GPT-5.5. In a short demo, the company showed the chatbot performing a live web search after a user asked when Argentina’s “next game” would be.

It understood the context and identified the next game as a 2026 World Cup quarterfinal match against Switzerland. It didn’t verbally say when the match would be, but the ChatGPT mobile app used during the demo showed additional information on screen.

Image credit: OpenAI

OpenAI also demonstrated how the model could do simultaneous translations, with ChatGPT translating English speech into Hindi in real time. Previous voice models could translate between languages but only sequentially, with one speaker talking first and ChatGPT translating afterward.

Paid users can choose from three levels of intelligence: instant, medium, and high. OpenAI said this lets users trade off speed against reasoning capability. The demo was shown using instant mode.