OpenAI launches GPT-Live-1 for full-duplex voice agents

OpenAI launched GPT-Live-1 in its API on September 10. The full-duplex voice model can listen and speak at the same time, handle interruptions, and connect to applications or phone systems. Voice sessions cost $0.05 per minute, billed by the second, with backend model and tool use charged separately.

For live conversations, GPT-Live-1 is designed to replace the handoffs among speech-to-text, reasoning and text-to-speech components in cascaded voice systems. Developers can still route deeper work to a separately chosen backend model or agent, using either OpenAI-managed Responses delegation or their own client, according to the company’s implementation guide. The application controls backend execution, context and which results return to the voice layer.

Several production responsibilities remain outside the model. OpenAI’s documentation says the developer’s application is responsible for permissions, user confirmations, private function execution and durable task state. Interrupting the agent’s speech does not automatically cancel backend work already in progress.

Developers can use a session prompt to set the agent’s role, tone, speaking pace, backchannel behavior and interruption policy, along with the conditions for delegating work. OpenAI’s prompting guidance says detailed workflows and tool checks should stay in the backend and application instead of being packed into the voice prompt.

GPT-Live-1 supports WebRTC, WebSocket, server-side sideband connections and SIP telephony. OpenAI also offers integration guidance for LiveKit, Twilio, Telnyx and Daily/Pipecat. The model returns automatic speech-recognition transcripts and response text, supports keyword biasing and alphanumeric understanding, and retains explicit turn detection even though it is not a turn-based model.

OpenAI introduced 12 named voice options for GPT-Live-1, which it said span more accents, dialects and languages. Access to custom voices requires eligibility and a request through the company’s sales channel.

OpenAI also reported a 30-percentage-point improvement over GPT-Realtime-2.1 on Full Duplex Bench. It said GPT-Live-1 ranked first on Tau3 when paired with GPT-6 Astra at medium reasoning effort. The launch materials did not include an independent benchmark report, leaving both results as vendor claims.

The early customer figures are also company-supplied. Speak said its evaluations found that GPT-Live-1 reduced interruptions during learner thinking pauses by almost 80% compared with its previous turn-based systems. Presence said its implementation simplified its codebase by 80% and removed 23,000 lines compared with its cascaded build. The launch page did not disclose enough methodology or sample information to validate either result independently.