OpenAI launches GPTLive voice model with full-duplex chat

OpenAI has launched a new generation of voice models called GPTLive, designed to make conversations with its ChatGPT assistant feel more natural and fluid. The models, which began rolling out globally on July 8, 2026, feature a "full-duplex" architecture that allows them to listen and speak simultaneously, enabling users to interrupt, pause, or ask the AI to slow down.

The update introduces two versions: GPTLive1 for paying subscribers on Go, Plus, and Pro plans, and a smaller, more efficient GPTLive1 mini for free users. Both are available on iOS, Android, and the ChatGPT website. A key technical shift is the model's ability to delegate complex tasks. When a user's question requires a web search, deeper reasoning, or other advanced capabilities, GPTLive can hand it off to OpenAI's latest frontier text model-currently GPT-5.5-and continue the spoken conversation while the background model works. The result is then woven back into the dialogue.

Company officials said this architecture addresses major limitations of previous voice modes, which operated on a rigid turn-based system. Those older models had to wait for a user to stop speaking before responding, often leading to...