OpenAI has unveiled its next-generation voice model, GPT-Live. Designed to enable natural, human-like interaction, GPT-Live serves as a new foundation model that replaces the existing ChatGPT voice functionality. OpenAI stated that GPT-Live will make voice conversations with AI significantly more natural and fluid.

The most significant feature is its full-duplex architecture. Previous voice AI relied on a 'turn-based' approach, where the system generated a response only after the user finished speaking. In contrast, GPT-Live can process user input continuously while simultaneously generating audio. This allows for a more human-like experience, enabling the AI to provide natural backchannel responses like "Hmm," "Yeah," or "I see," and to wait patiently without interrupting when the user pauses. It also supports real-time interpretation and translation.

To achieve this, OpenAI implemented a new architecture that separates voice interaction from complex processing. While GPT-Live handles the conversation itself, it delegates tasks requiring web searches, complex reasoning, or agentic actions to the latest large language model running in the background. Currently, GPT-5.5 performs this role, and the system is designed to automatically utilize new frontier models as they are released. This approach allows GPT-Live to maintain a natural conversation even while performing searches or reasoning tasks.

Performance has also seen significant improvements. OpenAI reported that GPT-Live-1 and GPT-Live-1 Mini outperformed the previous Advanced Voice Mode in overall conversation satisfaction, naturalness, flow, interruption frequency, and turn-taking. The company also noted that the new models surpassed their predecessors in internal benchmarks, including GPQA (which evaluates expert-level scientific reasoning), BrowseComp (which tests web navigation capabilities), and internal voice agent benchmarks for multi-turn communication tasks.

A new ChatGPT voice experience is also being introduced. Users can interrupt the AI at any time to ask questions, take a moment to think, or ask ChatGPT to speak more slowly. GPT-Live signals that it understands the user with natural responses like "Mhmm" or "Got it," and the existing nine voice options have been recalibrated to match the new model.

GPT-Live also allows users to select response modes based on the situation: 'Instant' for quick replies, and 'Medium' or 'High' reasoning modes for deeper thought. For specific queries regarding weather, sports, or stocks, the AI can display visual aids, providing a multimodal experience that combines voice and screen content. ChatGPT's memory and web search features are also fully integrated into voice conversations.

OpenAI emphasized that it has invested heavily in safety. GPT-Live incorporates a response framework reviewed by crisis response experts and age-appropriate policies for younger users. Various safeguards have been added to ensure that natural conversation does not inadvertently create risky situations. Extensive safety evaluations were conducted on voice output, and the model's ability to handle inappropriate requests has been improved.

GPT-Live will be rolled out sequentially to ChatGPT users worldwide starting July 8. Paid subscribers will have access to GPT-Live-1, while free users will have GPT-Live-1 Mini as their default voice model; both are available on iOS, Android, and the web. While video and screen-sharing features will not be available at launch, OpenAI plans to add them in a future update. The company also intends to provide GPT-Live via API in the future, enabling developers and enterprises to integrate this new real-time voice AI into their own applications.