What actually happened
On 10 September 2026 OpenAI put GPT-Live-1 in the API. Previously a voice bot waited for you to finish, turned speech into text, waited on a model, then spoke. Interrupt that loop and it broke. Now one model listens and talks at the same time, handles pauses and “actually, wait,” and can pass the hard work to another model or tool while the call keeps moving.
The voice layer is $0.05 a minute, billed by the second, not rounded up. Backend model and tool use is extra. An hour of talk time is $3 before reasoning, telephony and software. OpenAI says Full Duplex Bench rose 30 points versus GPT-Realtime-2.1. Speak told OpenAI unwanted interruptions during thinking pauses fell by almost 80% against its old turn-based stack. Yelp is using it on restaurant reservation and food-order calls and says callers now speak in fuller sentences. Those are customer reports, not your callers.
Why a small team cares
Take a dental clinic. A patient calls: “I need to move Thursday. Anything Friday afternoon?” GPT-Live-1 holds the conversation. A backend model checks the diary and comes back with slots. The patient cuts in: “Actually, Monday morning?” The voice keeps going while the diary is queried again. Once they pick a time, the booking tool writes it and the agent confirms.
The useful part is not a synthetic receptionist. It is the calls that already burn staff time because someone has to look something up or do a predictable action. Five hundred five-minute calls is 2,500 minutes. The GPT-Live-1 slice is about $125. Telephony, the backend model and the booking software sit on top. The number that matters is cost per resolved conversation, not cost per minute.
One healthcare customer told OpenAI the single-model voice layer cut their cascaded stack by 80% and 23,000 lines. Treat that as their build, not a promise for yours. For a shop hiring an agency, less glue still means less to maintain.
Hype vs useful
The hype is an agent that can discuss everything the business does. That is how you get wrong bookings and angry callers.
Useful is one job with a fence: reschedule, order status, or qualify a trade enquiry and create a callback. Define what the agent can read, what it can write, and when it must transfer. Measure resolved, escalated, abandoned and wrong. A natural voice makes people forget they are talking to software tied to your systems, so set the data rules before the first live hour.