OpenAI released GPT-Live, a new generation of voice models designed to make talking with ChatGPT feel closer to a real conversation. The key change is architectural: GPT-Live uses what OpenAI calls a full-duplex design, meaning it listens and speaks at the same time rather than waiting for a clean pause to respond.
It can interject brief acknowledgments like “mhmm” or “got it,” let a user interrupt naturally, stay quiet when asked to just listen, and hold its focus in noisy environments. Two versions are rolling out globally across iOS, Android and the web: GPT-Live-1 becomes the default for paid Go, Plus and Pro users, while GPT-Live-1 mini replaces Advanced Voice Mode for free accounts. An API version is planned soon.
The second design choice is delegation. GPT-Live handles the flow of conversation itself, but when a question needs web search, deeper reasoning or more complex work, it hands the task to a separate frontier model, GPT-5.5, and keeps chatting while the answer is prepared. This split lets OpenAI pair natural, low-latency speech with the intelligence of its larger models, and means GPT-Live can inherit smarter reasoning as OpenAI upgrades the model behind it.
Users can also pick a reasoning level, choosing Instant for speed or Medium and High for harder questions, and the interface can now surface visual cards for weather, stocks and sports while the conversation continues.
OpenAI frames this as the third generation of ChatGPT voice in two years, moving from the original stitched-together pipeline, to the smoother but still turn-based Advanced Voice Mode, to a system that makes interaction decisions many times a second. The company says more than 150 million people use ChatGPT’s voice and dictation features weekly, and that GPT-Live-1 was strongly preferred over the previous mode in its own head-to-head human tests.
Those comparisons are self-reported, and the launch has real limits: at launch GPT-Live does not support video or screen sharing in ChatGPT, cannot access memory in real-time mode, and may sound less fluent in some languages.
A Crowded Field
GPT-Live is less a breakthrough than OpenAI catching up to and integrating a capability that is quickly becoming standard. Google’s Gemini Live already offers full-duplex conversation plus camera and screen sharing, features GPT-Live lacks, while ByteDance shipped a production full-duplex system in its Doubao app and Nvidia released a customizable-voice model earlier this year.
OpenAI’s edge is its vast user base and tight coupling to GPT-5.5, not a technical monopoly on natural voice, which has already closed. Early reactions have been mixed in a telling way: some users find the frequent “mhmm” backchanneling reassuring, while others call it verbose and grating, a reminder that with voice, unlike text, the interface intrudes directly on attention. Whether GPT-Live’s feel wins people over, or wears thin, will be its real test.
