OpenAI released GPT-Live-1 in its developer interface. The model listens and speaks at the same time, rather than taking strict turns. It first appeared in ChatGPT before this release. Developers can pair it with a separate text model for deeper reasoning and tool use. They can also shape tone, pace and conversational style through instructions. The model handles background noise, pauses and interruptions without narrating each step. It supports phone calls, so agents can take reservations or customer support. It provides transcripts, keyword biasing and optional turn detection. OpenAI expanded the voice selection across accents, dialects and languages. Custom voices require a sales request.
What changed
Voice agents chained separate speech-to-text, reasoning and text-to-speech steps.
What it unlocks
Building phone-based voice agents that listen and speak at the same time.
- $0.05 per minute, voice layer
- +30 points on Full Duplex Bench
- ~80% fewer interruptions (Speak)
- 80% smaller code base (one customer)
What you need to act on it
- API access
- contact sales for custom voices
Sources