OpenAI launches GPT-Live voice models for natural AI conversations

OpenAI has released a new generation of voice models designed to make AI conversations feel more natural and responsive, according to details first reported by Axios. The GPT-Live-1 and GPT-Live-1 mini models represent the company's most sophisticated voice capabilities to date and will be available to free ChatGPT users starting this week.

The new models address longstanding friction points in voice-based AI interaction. Users can now interrupt the AI naturally mid-response, and the system won't prematurely cut off speakers who pause while formulating their thoughts. The models can also listen and speak simultaneously, which OpenAI says improves real-time language translation.

Why it matters

OpenAI's decision to make advanced voice capabilities available to free users signals a strategic bet that audio will become the dominant way people interact with AI systems. This shift has implications for everything from accessibility and workplace productivity to the design of future AI-powered devices and services.

Routing to frontier models

A significant technical upgrade involves how voice queries are processed. The new voice models now route user requests to OpenAI's most capable text-based reasoning models—addressing what company executives acknowledged has been a limitation for users who previously received lower-quality responses through voice compared to text chat.

Atty Eleti, product lead for ChatGPT Voice, described the release as just the beginning. The company envisions voice becoming "the primary interface to computing" over time, with applications extending to "increasingly complex, long-running, agentic work."

Economics and hardware implications

Voice interaction typically consumes more computational resources and tokens than text-based queries, making it more expensive to provide. OpenAI stated that declining inference costs have made the economics viable enough to extend voice access to free users—a notable shift in the company's service model.

The emphasis on voice also aligns with OpenAI's hardware ambitions. While the company declined to announce any hardware products during a press briefing, the focus on voice-first interaction could indicate plans for dedicated devices designed primarily for audio-based AI engagement.

Privacy and data handling

Audio data is retained for 30 days to maintain conversation context and memory features, though users can delete or export their data at any time. OpenAI has set training opt-out as the default for voice interactions, meaning conversations will only be used to train future models if users explicitly consent.

The details were first reported by Axios.