GPT-Live-1 is in the API: full-duplex voice agents at $0.05/min
OpenAI's model listens while it speaks, handles interruptions, and runs over phone lines — retiring the STT-LLM-TTS pipeline, with reasoning billed on top.
Copy markdown
Live now: $0.05 a minute for the voice layer
GPT-Live-1 is available in the OpenAI API. That $0.05/min buys only the front-end voice layer — the part that listens and speaks; the reasoning model and any tool calls you wire up bill separately, so budget the whole stack, not just the voice.
Full-duplex: it listens while it talks
Unlike turn-based voice, GPT-Live-1 handles interruptions, back-channels ('mhm'), pauses and background speech natively — one early eval reports roughly 80% fewer bad interruptions than prior turn-taking systems. For your users, that's a call that feels like a conversation, not a walkie-talkie.
The cascaded STT-LLM-TTS pipeline is retired
The old speech-to-text to LLM to text-to-speech chain is gone: a conversation layer runs concurrently while a separate backend model does the reasoning and tool execution. You keep your choice of brain — OpenAI reports #1 on Tau3 customer-service when paired with GPT-6 Astra — while GPT-Live-1 owns the ears and mouth.
Build this weekend: a phone agent
Telephony is a first-class target, so a voice receptionist, an outbound-call agent, or a hands-free assistant becomes a weekend build instead of a pipeline project. GPT-Live-1 posts a 30-point jump on Full Duplex Bench over GPT-Realtime-2.1; custom voices remain restricted for now.
The trade-offs before you commit
The $0.05/min is the voice layer alone, so a busy phone line adds up fast, and OpenAI hosts the loop — there's no self-host path. Real latency and cost ride on whichever backend brain you pair it with. Prototype the happy path, then price the full stack at your actual call volume.