OpenAI has introduced GPT-Live, a new generation of voice models designed to make conversations with ChatGPT feel closer to speaking with another person. Rolling out globally from today, GPT-Live replaces the traditional turn-based interaction with a full-duplex architecture that allows the model to listen and speak at the same time, reducing interruptions and making conversations more fluid.
The launch includes two models, GPT-Live-1 and GPT-Live-1 mini. GPT-Live-1 will power ChatGPT Voice for Go, Plus, and Pro users, while GPT-Live-1 mini will become the default experience for free users. OpenAI said the models will also be made available through its API in the future.
From Turn-Taking to Continuous Conversation
Previous versions of ChatGPT Voice relied on turn-based interactions, waiting for users to finish speaking before generating a response. GPT-Live changes that by continuously processing incoming audio while generating speech.
According to OpenAI, the new architecture allows the model to decide in real time whether it should respond, continue listening, pause, or interrupt naturally. That means users can pause to think, interrupt the assistant midway through a response, or ask follow-up questions without restarting the conversation. The company also says the approach improves live translation and creates a more natural conversational rhythm.
For users, the changes are designed to make everyday interactions feel less like issuing commands to a voice assistant and more like having a conversation. Whether someone is asking for directions while walking, practising a new language, or discussing a work project, the assistant can continue the interaction while carrying out more complex tasks in the background.
Smarter Responses Without Breaking the Flow
OpenAI has also separated conversation from reasoning. GPT-Live handles the real-time voice interaction, while more demanding requests such as web search or deeper reasoning are delegated to another model running behind the scenes.
At launch, GPT-Live uses GPT-5.5 in the background for those tasks. OpenAI said this approach allows the voice conversation to continue naturally while more complex processing happens without forcing users to wait in silence. The company added that future frontier models will be integrated into GPT-Live as they become available.
The updated ChatGPT Voice experience also introduces richer visual responses. During conversations, users may see cards for weather, sports, stocks, and other information while continuing to interact by voice.
Safety Remains Part of the Rollout
Alongside the new interaction model, OpenAI said GPT-Live includes voice-specific safety measures. These include safeguards for conversations involving self-harm, age-appropriate behaviour for teen users, protections against voice impersonation, and monitoring designed to reduce emotionally dependent interactions with AI. The company said it expanded its safety testing to include audio-native evaluations and internal red-teaming focused on voice-specific risks.
OpenAI also reported that GPT-Live performed better than or comparably to Advanced Voice Mode in its internal evaluations across conversational flow and several benchmark tests. Those evaluation results were conducted by the company and should be viewed as vendor-published performance data until independently validated.
Voice has become one of the fastest-growing ways people interact with ChatGPT. OpenAI says more than 150 million users engage with ChatGPT through voice and dictation every week. Rather than treating voice as a speech interface layered on top of a language model, GPT-Live reflects a broader shift toward AI systems designed to maintain conversations while reasoning, searching, and completing tasks simultaneously.
For enterprises and developers building voice-enabled applications, the update also signals a move beyond simple question-and-answer interactions toward assistants that can manage longer, more dynamic conversations without sacrificing access to more capable reasoning models.
