Updated September 14, 2026
AI + Business
Free · Independent · Grant-funded
AICV News
News · Views · Field Notes

Field Notes

openai's phone-ready voice model listens and talks at once, at $0.05 a minute

The people who answer the phone at a restaurant, a hotel front desk or a clinic here are who this touches; the calls OpenAI names are reservations and customer support.

September 14, 2026 · · Coachella Valley

Source: OpenAI (article, September 10, 2026) · archived

A phone agent you can interrupt mid-sentence, that keeps its place and keeps talking, is now something a developer can rent by the minute. OpenAI put GPT-Live-1 into its API on September 10 at $0.05 a minute for the voice layer. What’s different is that one model listens and talks at the same time. OpenAI’s own description of the usual voice bot: it transcribes you, ships the text to a reasoning model, reads the answer back, and every handoff adds a pause. This one skips the handoffs.

Where the thinking happens is up to whoever builds it. The voice model passes hard questions to a backend model, and OpenAI’s example pairs it with a cheap model for scheduling and order updates and a strong one for the complicated calls. Tone, pace and personality come from the system prompt. It sits through background noise and silence without narrating, keeps its place across a long call, and takes phone calls. Yelp says Yelp Host is already answering restaurant calls with it, reservations and food orders.

The $0.05 buys the voice only. The model doing the thinking is billed separately, so the cost of a full call isn’t on the page. New voices across accents, dialects and languages ship with it; a custom voice goes through sales.

Not verified by us

  • Speak's 'almost 80%' fewer interruptions, Yelp's 'meaningful improvements in call handling rates' and a health-care customer's '80%' smaller code base are customer testimonials quoted by OpenAI, not independent measurements.
  • The 30-percentage-point Full Duplex Bench gain over GPT-Realtime-2.1 and the Tau3 ranking are OpenAI's own benchmark claims.
  • $0.05 a minute covers the voice layer only; the page does not price the backend model's tokens, so the cost of a full call is not stated.

How this was made: researched and written at the desk with AI, every figure checked against the source; edited independently by a second model; published by Sat Singh on September 14, 2026.