OpenAI makes GPT-Live-1 generally available in the API

OpenAI made GPT-Live-1 generally available in the API on September 10, 2026, a dedicated voice model that listens and speaks at the same time and hands reasoning and tool use off to a backend model, priced at $0.05 per minute for the front-end voice layer.

What's new

GPT-Live-1 was first introduced inside ChatGPT; the API release makes it available to any developer building voice products. OpenAI is positioning it as a separate "voice layer" that sits in front of a reasoning model rather than a single model doing both jobs at once:

• Full-duplex audio — the model can listen and speak simultaneously, rather than waiting for a turn to end before generating a response.
• Natural interruption handling — OpenAI says the model reduced unwanted interruptions by roughly 80% in its evaluations, letting users interject mid-response the way they would in a human conversation.
• Background noise management for real-world audio conditions rather than clean, single-speaker input.
• Long-session reliability, with improved context retention over extended calls.
• Telephony support, enabling full-duplex voice agents for phone calls — OpenAI cites use cases from restaurant reservations to customer support.
• Customizable tone, pace, and style via system prompts, and delegation of tool calling to whatever backend model it is paired with.

Pricing is $0.05 per minute for the front-end voice layer, billed separately from whatever backend reasoning model handles the actual task.

Context

OpenAI's own description draws a clean line between what GPT-Live-1 does and what it delegates: "GPT‑Live‑1 handles listening and speaking in a single model, simplifying the voice layer. It can respond to interruptions and acknowledgements as they happen, while delegating deeper reasoning to the back end." That split — a fast, always-listening voice front end paired with a slower, more capable reasoning model behind it — is the same architecture OpenAI has been building toward since its original realtime voice models, but GPT-Live-1 is framed as the general-availability, production-ready version of that layer.

The release arrives the same day OpenAI also took its new Agents API out of beta-only territory and into public beta, continuing a pattern this month of OpenAI unbundling capabilities that used to live only inside ChatGPT — voice, agent orchestration — into standalone, separately priced API primitives.

Why it matters

Full-duplex, low-latency voice is the hardest remaining gap between talking to an AI assistant and talking to a person, and it's also the layer most directly competing with Google's Gemini Live and other real-time voice offerings. By selling GPT-Live-1 as a thin, dedicated voice layer rather than folding voice into a single monolithic model, OpenAI is making a architectural bet that voice and reasoning should scale independently — developers can upgrade the reasoning model behind a voice agent without touching the voice layer itself. The $0.05-per-minute price, combined with telephony support, is also a direct pitch at commercial voice-agent use cases like phone-based customer support and reservations, markets currently served largely by specialized voice-AI startups rather than frontier labs.