Build More Natural Voice Experiences with GPT-Live-1 in the API

What changed

Source summary

OpenAI released GPT-Live-1, a new model available via the API for building real-time voice experiences, with smooth interruption handling and the ability to delegate reasoning and tool calls to a backend text model instead of chaining separate speech-to-text, LLM and text-to-speech systems. It is priced at $0.05 per minute for the voice layer.

Why it matters to your work

Practical interpretation

Relevant mainly if your organisation builds or commissions custom voice apps (phone assistants, in-app voice agents) — it simplifies the tech stack developers need and sets a clear per-minute cost, useful context if you're scoping or budgeting a voice-AI project.

When this could help

Building full-duplex voice agents for phone calls (reservations, support), or any app needing natural, interruption-tolerant spoken conversation

A single model reasons over incoming and outgoing audio together, handling natural interruptions and background noise without the latency and brittle handoffs of chaining separate speech-to-text, LLM, and text-to-speech systems, and can delegate deeper reasoning or tool calls to a backend text model

What to check

• Voice and language/accent selection is still expanding, with more promised over the coming months Source ↗
• Custom voice access is gated behind a sales eligibility process, not self-serve Source ↗
• Performance claims (e.g. 30-point Full Duplex Bench improvement, #1 on Tau3 paired with GPT-6 Astra, ~80% fewer interruptions) are OpenAI's own reported figures, not independently verified Source ↗