OpenAI's GPT-Live-1 brings full-duplex voice AI to the API at 5 cents a minute

What happened

OpenAI released GPT-Live-1 into its API on September 10, priced at $0.05 per minute for the voice layer. It's a full-duplex speech model: it listens and talks at the same time, the way a person on a phone call does, instead of taking turns in the stiff record-then-respond pattern most voice bots still use. Early integrations are already live — Yelp is using it for restaurant reservation calls, Fin for customer support, Speak for language tutoring, and Cognition has paired it with its AI engineer Devin.

What's genuinely new

Every voice bot you've dealt with so far, including most of the ones businesses have deployed, is three models stitched together: speech-to-text, a reasoning model, then text-to-speech, taking turns. That's why they interrupt badly, go silent while "thinking," and lose the thread the moment a caller talks over them. GPT-Live-1 collapses the audio side into one model that handles listening and speaking simultaneously, and hands the actual reasoning off to whatever backend model you pair it with.

OpenAI published benchmark numbers against its previous voice model, GPT-Realtime-2.1, and they're a real jump, not a rounding error: turn-taking latency drops from 1.41 seconds to 0.798 seconds, and on OpenAI's own Tau3 voice intelligence benchmark the pass rate goes from 45.7% to 86.2%. It also does native transcription, turn detection, keyword biasing, and holds context across long sessions — the stuff that makes a phone call feel coherent instead of like a form you're filling out verbally.

The pricing model matters as much as the capability. Five cents a minute buys you the voice layer only. The reasoning underneath — the part that actually decides what to say, checks a calendar, or looks up an order — is billed separately at standard token rates for whatever model you connect it to. That's a deliberate unbundling: cheap, fast audio in front, and you choose how much brain to put behind it.

What it means for a business owner

If you run anything with an inbound or outbound phone queue — reservations, appointment booking, order status, basic support triage — this is the first version of "AI answers the phone" that doesn't sound like it's fighting the caller. The latency and interruption numbers are the actual bottleneck that's kept voice automation out of production for most businesses; a bot that can't handle a caller talking over it gets escalated to a human within the first ten seconds, every time.

The unbundled pricing also changes how you'd scope a build. A simple booking or FAQ line can run on a cheap backend model and stay near that $0.05/minute floor. Something that needs to reason over your actual order history or make judgment calls costs more, but you're not paying premium rates for the audio itself — only for the intelligence you actually need. That makes it easier to price out a voice agent project against what a human answering the same volume of calls costs you today, since the audio cost is now a fixed, known number instead of bundled into an opaque per-minute rate.

The honest caveat

This is a component, not a finished product. OpenAI ships the voice layer; someone still has to build the logic that pairs with it — the calendar integration, the order lookup, the escalation path when the caller wants something the bot can't do. That's most of the actual engineering work, and it's exactly where voice-agent projects go wrong: not in how natural the voice sounds, but in what happens when the conversation goes off-script.

Custom voices still require going through OpenAI sales rather than self-serve, and broader language and voice coverage is described as coming in future months, so if your use case needs a specific accent, language, or brand voice today, check availability before you scope anything. And a benchmark win against OpenAI's own prior model tells you it's better than what OpenAI shipped before — it doesn't tell you how it holds up against a genuinely hostile caller, a bad phone connection, or someone speaking a language it wasn't tuned for.

What to do about it

If phone volume is a real cost center for you, don't start by picking a model. Start by pulling the actual numbers: call volume, average handle time, and what percentage of calls are simple enough to script versus what percentage need real judgment. That ratio tells you whether a voice agent pays for itself, and it's the same math regardless of which voice model ends up underneath it.