OpenAI launched GPT-Live, the new voice mode of ChatGPT, with global deployment starting today on iOS, Android, and ChatGPT.com. The change is not cosmetic: GPT-Live introduces an architecture called full-duplex that allows the AI to listen and generate responses simultaneously, eliminating the rigidity of the turn-taking that characterized previous versions.
To understand the scope of the change, one must look at how previous generations operated. The first voice mode of ChatGPT worked as a cascading system: one model transcribed the voice, another generated the response, and a third converted it back into audio. The advanced voice mode improved that experience by processing audio within a single model, but it still operated turn-wise, waiting for the user to finish speaking before responding. GPT-Live breaks that logic: the system makes decisions multiple times per second about whether to speak, listen, pause, or delegate a task to another model.
In practice, this translates to a different conversation. The user can interrupt with a question halfway through a response, pause to think without the system interpreting silence as the end of the conversation, or ask it to speak more slowly. The voice can emit brief signals like "mmm" or "yes" to indicate that it is still listening, and the system has improved its ability to distinguish the user's voice amid background noise. For quick queries like the weather, sports, or stock data, GPT-Live generates visual responses in the form of cards that appear on screen while it speaks.
When the conversation requires something more than an immediate response, GPT-Live can delegate web searches, extended reasoning, or complex tasks to the GPT-5.5 model in the background, while keeping the conversation active with the user. OpenAI noted that it will update the support model as it releases new generations. The mode will also allow users to choose between different levels of reasoning: Instant for quick answers, and Medium or High for tasks that require more depth.
The distribution by user type is clear. GPT-Live-1 will be the default model for Go, Plus, and Pro plans. Users with free accounts will access GPT-Live-1 mini, the more compact version. According to OpenAI's internal evaluations, both models outperform the previous advanced voice mode in conversations lasting five to ten minutes, both in fluency and perceived naturalness, handling interruptions, and turn-taking. The nine voices of ChatGPT were remastered for this version. The launch also includes support for search, memory, images, and file uploads from the voice mode.
There is a relevant limitation in this first version: GPT-Live does not support voice with video or screen sharing for now, although OpenAI confirmed it is working on incorporating those capabilities. API access also does not have a release date yet.
For the next+ team, GPT-Live is probably the most relevant change to the ChatGPT interface since the launch of GPT-4o. Not because it is smarter, but because it addresses the right problem: friction in conversation. An AI system that knows when to be silent, when to wait, and when to resume without the user having to mark explicit turns fundamentally changes the nature of interaction more profoundly than a benchmark improvement. For organizations evaluating the use of AI voice assistants in customer service, internal support, or workflow automation, the full-duplex architecture of GPT-Live addresses one of the most cited limitations for mass adoption: that conversation does not feel natural. When that friction disappears, the threshold for adoption decreases significantly.