The AI Voice Model That Listens While It Talks Just Hit the API: Retest the Call Yours Keeps Botching
OpenAI's full-duplex voice model went live in the API on September 10. It removes the single behavior that makes callers hang up on phone bots, and it runs about 20 cents for a four-minute booking call.
What shipped. OpenAI put GPT-Live-1 into its API on September 10, 2026. The change that matters is architectural: the model listens and speaks at the same time. Every voice agent before it waited for silence to decide you were finished talking.
Why silence was the problem. A customer pausing to read an address off a bill, a dog barking, a truck going by: a turn-based agent reads all of that as "done talking" and starts speaking over the caller. That one behavior is why most people hang up on a phone bot and dial a competitor. Speak, a language-learning app, reported cutting wrong interruptions by roughly 80% after moving to the new architecture.
What it costs. $0.05 per minute for the voice layer. A four-minute booking call runs about 20 cents. Two hundred calls a month is roughly $40 of voice, plus whatever your platform charges on top of it.
The move this week.
- Pick the one call type your agent handles worst. Usually after-hours booking, or the caller who has to go look something up mid-call.
- Check whether your voice platform has exposed the new model yet. Most wrap the OpenAI voice API within a few weeks of release.
- Run five test calls yourself. On three of them, go quiet for four full seconds mid-sentence, the way a real customer does when they are digging for a policy number.
- Score exactly one thing: how many times it talked over you.
- If interruptions drop to zero or one, route that call type to the new model and keep your existing script. The script was probably never the problem.
Keep your data straight. A voice agent that reads from your customer list is touching real records. Run it on a business account with training turned off, and give it a scoped login instead of an admin one.