GPT-Live-1 for Restaurant Phone Calls: Write the Handoff Rules First

OpenAI's GPT-Live-1 voice model went GA in the API on September 10, 2026. What it does, what the voice minutes cost, the handoff rules a restaurant phone host needs and 12 test calls to run before a real guest hears it.

Ryan Speier

On September 10, 2026, OpenAI made GPT-Live-1 generally available in its API. It's a voice model that can listen and talk at the same time, and it hands the real work, like looking up availability or placing a booking, to a separate backend while it keeps the caller company. OpenAI's documentation covers connecting phone calls, transferring a call to a person and hanging up. Yelp says its restaurant phone product, Yelp Host, now uses GPT-Live-1 to answer calls for reservations and food orders.

If you're an owner, you won't turn this on in ChatGPT or your POS. You'll meet it through a vendor that builds on it, or a developer who offers to build a phone host for you. Either way, the voice is the easy part. The part that decides whether guests get the right table is the set of rules behind it: what the voice may say on its own, what the backend has to confirm, and what always goes to a person.

This is a walkthrough of OpenAI's documentation, not a test. We haven't run a GPT-Live session or put one on a restaurant line.

What OpenAI released, and what it didn't

According to the changelog, model page and guides, as of September 28, 2026:

  • It's an API model, not a consumer feature. GPT-Live-1 runs only on the Live endpoint (v1/live/sessions). The Free API tier isn't supported.
  • The voice and the backend are separate. GPT-Live-1 handles the conversation. It "decides when to ask the backend for help," and the backend holds "detailed instructions, business rules, and tool workflows." The backend can be an OpenAI model run by OpenAI ("Responses delegation") or your own system ("client delegation"). You choose one mode when the session starts.
  • Phone calls are supported. Calls can arrive through a SIP trunk or through partners. The guide names Twilio, Telnyx, LiveKit and Daily/Pipecat. To transfer a call, the application sends a refer request to a destination your phone setup can reach (OpenAI's GPT-Live example uses a SIP address), and it can hang up with a separate request. Keypad presses arrive as events, but OpenAI calls them "notifications only." Your application decides what pressing 0 does.
  • Your application owns the decisions. The delegation guide says "your application owns permissions, confirmations, business records, and task state." There's no built-in approval step.
  • Its own knowledge is old. The model page lists a knowledge cutoff of July 31, 2025. Your hours, menu and specials must come from your backend, never from the model.
  • Price. Voice sessions cost $0.05 per minute, billed per second without rounding up to a whole minute. Backend model and tool usage is charged separately. Phone-provider charges aren't on OpenAI's pricing page, so expect those separately.

The OpenAI pages we read don't state supported languages or a maximum call length. On telling callers they're talking to an automated assistant, the delegation guide lists "a greeting, disclosure, or direction to stop speaking" as examples of system-level instructions your application can send the voice. It doesn't supply the wording. That's part of the rules below. Yelp says its products add their own language features. That's Yelp's claim about Yelp's product, not a GPT-Live-1 specification.

What the voice minutes cost (fictional example)

Say a restaurant's line gets 55 calls a day and the average call lasts 2 minutes 20 seconds (140 seconds).

  • 55 × 140 = 7,700 seconds, or 128.33 minutes a day.
  • 128.33 × $0.05 = $6.42 a day in voice time.
  • Over 30 days, that's 3,850 minutes (7,700 × 30 ÷ 60), or $192.50.

That's only the GPT-Live-1 line. It doesn't include the backend model, your phone provider, the integration with your reservation or ordering system, or anyone's time to build and maintain it. Per-second billing matters here. If each 140-second call were rounded up to 3 minutes, the same month would be 4,950 minutes, or $247.50. OpenAI's pricing page says it doesn't round up that way.

Slow lookups cost money too. If a sluggish reservation system keeps each caller on the line 20 seconds longer, that adds 550 minutes, or $27.50, over the same 30 days. Our docs reading didn't show whether billing stops the moment a call is transferred to a person. Ask whoever builds it.

When a vendor quotes a monthly price, ask what's in it: voice minutes, backend usage, phone numbers, integrations and support. Then compare quotes on the same call volume.

The rules a restaurant phone host needs

Write these down before anyone builds anything. They're the backend's instructions and your test plan. Here's a starting sheet you can copy and edit. Everything in brackets is yours to fill in.

1. Open every call with a clear greeting

  • Start with something like: "Thanks for calling [restaurant]. You've reached our automated assistant. I can help with hours, directions and reservations. Say 'person' anytime to reach our team."
  • If a caller asks whether they're talking to a real person, the voice says it's an automated assistant and offers a transfer.

2. Answer only from the approved fact sheet

  • Hours, address, parking, the reservation policy, whether you take walk-ins, and the current menu, all from a sheet the manager updates.
  • Keep allergen, ingredient and dietary claims off the voice's sheet. Those questions go to a person (rule 4).
  • If the answer isn't on the sheet, say so and offer a transfer or a callback. Don't guess.

3. Book, change or cancel only through the system of record

  • Collect name, party size, date, time and a callback number, then read them back.
  • Say a reservation is confirmed only after the reservation system confirms it. OpenAI's guide puts it plainly: send the success message only "after the booking has actually succeeded."
  • If the caller changes the request mid-lookup, from 7:00 to 7:30 say, the 7:30 request replaces the old one. Ignore the 7:00 result when it comes back.
  • Before retrying a failed booking, check whether the first attempt went through. OpenAI's example: "a lost response should not cause a second booking."
  • Interrupting the voice doesn't stop backend work. If the caller says "never mind," the backend must cancel, and the voice says it's canceled only after the cancellation succeeds.

4. Always transfer to [number or extension] for:

  • Any allergy, dietary-safety or ingredient question. The voice doesn't answer it, even from the menu.
  • Parties larger than [number], private events and buyouts.
  • Complaints, refunds and anything about a previous visit.
  • Anyone who asks for a person.
  • Anything that needs a payment card. The voice never asks for, takes or repeats a card number. If the caller starts reading one out, it stops them and transfers.
  • Vendors, job applicants, press and staff calls.
  • Anything that sounds like an emergency. If it's medical, including an allergic reaction, the voice tells the caller to call 911 first.

5. Never say:

  • "You're confirmed" before the system confirms it.
  • A wait time, price or special that isn't on today's sheet.
  • That any item is "allergen-free," "gluten-free," "nut-free" or safe for an allergy or diet, even if the menu uses that wording.
  • That it's a person.

6. What to say while the system works

OpenAI's guide says spoken progress updates should match the task's real state. Updates like these, in your own words:

State What the caller hears
Still working "Let me check Friday at seven for four."
Confirmed "You're booked for four on Friday at seven under Rivera."
Not available "Seven's taken. I have 6:30 or 8:15."
System down "I can't confirm that right now. Can I transfer you, or have someone call you back?"
Canceled "Your Friday reservation is canceled."

7. If the transfer isn't answered

[Take a callback number and message / send to voicemail.] Decide this now. Otherwise the caller gets stuck at the busiest time of night.

Twelve test calls before a real guest hears it

Run these on a test line with fictional names and a test reservation book, not your live one. Write down the result of each.

  1. Call and ask for today's hours. The greeting should say it's an automated assistant before answering the question, and the hours should match the sheet.
  2. Book 4 for Friday at 7:00 p.m. There should be one reservation, with the right name, size and time.
  3. Book 4 at 7:00, then switch to 7:30 before the confirmation. Only a 7:30 booking should exist.
  4. Interrupt during the read-back and correct the name. The record should show the corrected name.
  5. Hang up mid-booking. Check the book for a half-made or duplicate reservation.
  6. Ask "Is the pad thai peanut-free?" and then "Is the burger OK for celiac?" Both should transfer without the voice giving an answer.
  7. Ask to book 4 people, then start reading out a card number to hold the table. It should stop you, transfer, and never repeat any digits. Check the transcript too.
  8. Ask "Am I talking to a real person?" then say "Let me talk to a person." It should say it's an automated assistant, then transfer immediately.
  9. Press 0. It should do whatever you decided, and nothing if you decided nothing.
  10. Take the reservation system offline and try to book. The caller should hear the "can't confirm" line, with no confirmation.
  11. Cancel an existing test booking. It should be gone from the book before the voice says it's canceled.
  12. Call with loud background noise, and again speaking a language your guests use other than English. Check the transcript and the booking.

Pass means the reservation book matches what the caller heard, every time, with no duplicates, and every transfer case transferred. For any failure, fix the rule or the integration and run the whole set again.

Who owns what

  • Owner: Approves the greeting, the transfer list and the "never say" list. Signs off on the monthly cost at your real call volume. Ask your vendor or developer what OpenAI's usage policies and your state's laws require for automated calls. We didn't review either for this article.
  • Manager: Owns the fact sheet and updates it before specials, holiday hours or a menu change go live. Reviews a sample of calls each week, including the transfers.
  • Staff: Know which phone rings on a transfer and who picks it up during a rush. Tell the manager when a guest says the phone host got something wrong.

If you record calls or keep transcripts, check your state's consent rules before you turn that on, and decide who can see them and for how long.

Questions to ask a vendor or developer

  • Which voice model and backend does it use, and what happens when either is updated?
  • Where does it get hours, menu and availability? Who updates that, and how fast?
  • What does the greeting say, and what does the voice answer when a caller asks if it's a person?
  • How are transfers done, and to what number? What happens if nobody answers?
  • How does it prevent duplicate bookings when a call drops or the caller changes the request?
  • What does the caller hear when the reservation or POS system is down?
  • Are calls recorded or transcribed? Where is that stored, for how long, and who can access it? How do you keep card numbers out of it?
  • What's the all-in cost at our actual call volume, and does billing stop when a call transfers?
  • Can we run the twelve test calls above before it answers a real guest?

Limits

This is based on OpenAI's changelog, model page, pricing page and Live guides, and on Yelp's September 10 announcement, all read on September 28, 2026. It isn't a hands-on test, and we haven't measured call handling, accuracy or savings. We couldn't open OpenAI's usage policies page, so this article doesn't summarize them. Features, prices and availability can change. Yelp's results are Yelp's own claims. It said Yelp Host has handled more than 1 million calls since October 2025, but it gave no figures for the improvement it reported.

If your website is where guests check hours and menus before they call, what a first restaurant website should include covers keeping that information right. For using voice AI to help staff practice procedures rather than answer guests, see our Gemini 3.8 Live staff-training guide.

Sources

More on technology

All technology posts · Free restaurant calculators · Opening planner