GPT-Live-1 API: $0.05/min Full-Duplex Voice, Pricing & Setup

GPT-Live-1 is OpenAI's full-duplex voice model, released to the API on September 10, 2026. It can listen while speaking, reacting to interruptions and backchannel cues the way a human conversation does. It runs exclusively on a dedicated Live endpoint (v1/live/sessions) — a separate service from the Realtime API, Chat Completions, or Responses API. Pricing is $0.05/minute, billed per second, with no free tier. In short: for use cases where natural, interruption-tolerant voice conversation matters — phone support, voice reception, live agents — GPT-Live-1 is currently the strongest option, especially where tool-calling accuracy matters too.

Input and output are limited to audio and text; there is no image or video support. The knowledge cutoff is July 31, 2025. It supports streaming responses, function calling, and delegating work to a backend, but does not support structured outputs or fine-tuning. It runs on the same underlying model that powers ChatGPT Voice, and Yelp has already deployed it for phone reservations.

- 12 new voices spanning a wider range of accents, dialects, and languages
- Full-duplex (simultaneous two-way) audio lets it react to interruptions instantly
- Function calling and backend task delegation are supported natively
- If a user interrupts, delegated backend work is not automatically cancelled — task state persists

Billing is $0.05 per minute, calculated per second. This covers the Live model's own audio processing; token costs for whichever backend model it calls, plus any tool usage, are billed separately. There is no free tier.

Cost example: a 3-minute call costs $0.15 for the voice portion alone ($0.05 x 3 min), with backend token costs adding a few more cents to tens of cents. For a call center handling 1,000 calls a month at an average of 3 minutes each, the voice portion alone works out to roughly $150/month ($0.05 x 3 min x 1,000 calls), with backend and tool usage billed on top.

1. Start a session on the dedicated Live endpoint, v1/live/sessions — distinct from the Realtime API or Responses API endpoints
2. Pick a connection method — WebRTC for browsers, WebSockets for server-to-server, or telephony/SIP for phone agents
3. Choose how backend work is delegated — either OpenAI's built-in managed Responses delegation, or client-side delegation to your own models, agents, or services
4. Design around durable task state, since a user's interruption does not automatically cancel in-flight backend work
5. Optionally pair it with the Codex SDK to pass conversation context directly into a Codex thread, letting a voice instruction continue as a development task

For how to design the backend delegation and session management, see our OpenAI Agents API guide, which covers sessions and sandboxing in more depth.

Against GPT-Realtime-2.1, GPT-Live-1's published benchmarks show full-duplex interactivity at 80.1% vs 45.4%, turn-taking latency of 0.8s vs 1.4s, tool-calling accuracy of 87% vs 60%, and a banking voice support pass rate of 32% vs 12.4%. Paired with GPT-6 Astra, it reaches 83.6% Tau3 task completion and 90% response quality. Compared to the earlier GPT-Realtime-2 models, full-duplex support and interruption handling are now the clear differentiators. If you need to keep data on-premises or avoid per-minute API billing, our NVIDIA PersonaPlex guide covers a local, open-source alternative; if you want a managed service with the latest benchmark performance, GPT-Live-1 is the stronger pick.

- Phone reservations and reception: connect over telephony/SIP to build a phone agent; interruption tolerance matters most here (Yelp's phone reservation use is a live example)
- First-line customer support: suited to routine intake; however, the banking voice-support pass rate is still only 32% (vs 12.4% before), so design complex or high-risk cases around human escalation
- Development assistant agents: pairing with the Codex SDK lets a spoken instruction carry straight into a development task
- Cautions: no image or video input, so it is unsuited to conversations that need shared visual material; the July 31, 2025 knowledge cutoff means current information needs to come through tool calls or search; with no free tier and per-second billing, designs with long idle time can get expensive