OpenAI Launches GPT-Live-1, a Full-Duplex Voice Model That Cuts Interruptions by 80%
OpenAI launched GPT-Live-1, a voice model built for natural, full-duplex conversation β listening and speaking at the same time in a single unified architecture rather than chaining separate speech-to-text, language model, and text-to-speech components.
Details
- Interruption handling: cuts interruptions by nearly 80% compared to turn-based voice systems by reasoning over audio simultaneously rather than waiting for a full turn to end
- Conversational naturalness: manages background noise, silence, and side conversations without breaking the flow of dialogue
- Tool delegation: routes complex reasoning tasks to backend models like GPT-6 Astra while keeping the live conversation continuous
- Customization: developers can shape tone, pace, and conversational style through the system prompt, and eligible customers can request custom voices
- Benchmarks: improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, and ranks #1 on Tau3 when paired with GPT-6 Astra for frontier voice-agent intelligence
- Telephony-ready: supports full-duplex voice agents for phone-based applications, with demonstrated use cases including restaurant reservations, customer support, language tutoring, and banking with knowledge retrieval and account tools
- Pricing and availability: $0.05 per minute for the front-end voice layer, available immediately in the API
What happened next
GPT-Live-1 gives developers a voice front-end explicitly designed to pair with reasoning models like GPT-6 Astra rather than working as a standalone voice bot, positioning OpenAIβs voice stack around a division of labor between fast conversational handling and deeper backend reasoning.