OpenAI has introduced GPT-Live, a new family of voice models designed to make conversations with ChatGPT feel more natural by allowing the AI to listen and speak simultaneously. The company is rolling out GPT-Live-1 and GPT-Live-1 mini to ChatGPT users globally today, with GPT-Live also set to arrive in the API at a later date.
The new models replace the traditional turn-based approach used in previous ChatGPT voice experiences with a full-duplex architecture that continuously processes incoming audio while generating speech. Rather than waiting for a user to finish speaking before responding, GPT-Live can acknowledge what it hears with brief responses, remain silent when appropriate, or naturally interrupt and resume conversation as needed. According to OpenAI, the system makes interaction decisions many times per second, including whether to continue listening, pause, speak, interrupt, or invoke a tool.
OpenAI also separated conversational interaction from more computationally intensive tasks. When a request requires web search, deeper reasoning, or more advanced capabilities, GPT-Live hands that work to a frontier model running in the background while continuing the conversation. At launch, GPT-Live delegates those tasks to GPT-5.5, and OpenAI said the underlying reasoning model will be updated as newer frontier models become available.
The architecture is intended to improve responsiveness without sacrificing access to more capable AI models. While background tasks are running, users can continue speaking instead of waiting for the system to complete its reasoning before responding.
GPT-Live powers a redesigned ChatGPT Voice experience that introduces several new behaviors. Users can interrupt the assistant while it is speaking, pause without being cut off, or ask ChatGPT to remain quiet while listening. OpenAI said the system is also better at filtering background noise and maintaining focus on the speaker’s voice. The company has also remastered ChatGPT’s nine built-in voices for the new models.
Voice conversations can now display visual cards alongside spoken responses for subjects such as weather, stocks, and sports. Existing ChatGPT Voice capabilities, including search, memory, image support, and file uploads, remain available.
OpenAI said users can choose different reasoning levels depending on the task, with Instant prioritizing speed and Medium or High allocating more time for reasoning.
The company also published evaluation results comparing GPT-Live with Advanced Voice Mode. OpenAI said GPT-Live-1 and GPT-Live-1 mini were preferred in human evaluations covering conversational flow, turn-taking, interruptions, overall preference, and naturalness during five- to ten-minute conversations.
Beyond conversational quality, OpenAI reported improvements across several benchmarks. GPT-Live-1 outperformed Advanced Voice Mode on GPQA, which measures scientific reasoning, BrowseComp, which evaluates agentic web search, and an internal telecom support benchmark designed for full-duplex voice agents.
According to OpenAI, more than 150 million people use ChatGPT’s Voice and Dictation features each week for tasks ranging from hands-free assistance to language practice and storytelling. GPT-Live is intended to serve as the foundation for future voice experiences that support increasingly complex and longer-running AI interactions.
The company said safety was a core focus during development. GPT-Live includes additional audio-native safety testing, synthetic voice evaluations, and internal red-team exercises covering areas such as self-harm, emotional reliance, violence, sexual content, psychosis, and mania. OpenAI said the model performed comparably to or better than Advanced Voice Mode across nearly all evaluated categories.
The system also includes safeguards that operate during live conversations. When potentially unsafe responses are detected, GPT-Live can redirect the conversation, provide additional safety guidance, or end the interaction in higher-risk situations. OpenAI said it has adapted its existing self-harm support flows for voice conversations and added age-appropriate behavior for teen users, with parental controls available for ChatGPT Voice.
OpenAI emphasized that GPT-Live is designed for conversational interaction rather than voice impersonation. The system uses a predefined set of voices and includes protections intended to prevent it from imitating the voice of a real person.
GPT-Live is rolling out across iOS, Android, and ChatGPT on the web. GPT-Live-1 becomes the default voice model for Go, Plus, and Pro subscribers, while GPT-Live-1 mini is the default for Free users. At launch, GPT-Live does not support voice conversations with video or screen sharing, although OpenAI said those capabilities are planned for a future release. Legacy versions of ChatGPT Voice will continue to offer those features until they are added to GPT-Live.
About this article: This article was generated with AI assistance and reviewed by our editorial team to ensure it follows our editorial standards for accuracy and independence. We maintain strict fact-checking protocols and cite all sources.
Word count: 729Reading time: 0 minutes
Explore More AI Resources
Continue with high-value guides related to this topic.
