GPT-Live voice agent: create a voice from one sentence and act while you talk

A GPT-Live voice agent can now be built with a voice of its own, created from a single sentence, and with actions that run while the customer is still talking. On October 15, 2026, OpenAI built Decisions natively into GPT-Live, and on October 16 it opened voice creation from a text description to every API customer. Creating the voice is free; speaking with it is billed like any other GPT-Live session.

What changed in GPT-Live this week

The news is not "a new voice": it is two changes that arrived a day apart. GPT-Live is OpenAI's voice model that listens and speaks at the same time (full duplex), and it has been in the API since September 10. This week's changes make it usable for a business: a voice of your own, and an agent that also does things. According to the Invoca Buyer Experience Benchmark, 68% of 500 US consumers prefer calling a business over chat or email, so voice still matters.

Photo: Shixart1985, CC BY 2.0, via Wikimedia Commons.

How a GPT-Live voice agent works: two brains

Think of two brains. GPT-Live holds the conversation: it listens while it speaks, you can interrupt it, and it does not lose the thread. Behind it sits another model that reasons and touches your systems. OpenAI calls this delegation, and there are two kinds: Responses delegation, where OpenAI runs the model for you, and client delegation, where you connect your own and control the context.

The key piece this week is Decisions. Instead of answering in free text, it answers closed-choice questions: raise the temperature or nothing, turn on the defogger or nothing. A system that answers with closed choices is one your code can execute without guessing.

The full walkthrough, in six steps

1. Create the voice. One request to /v1/audio/voices with type: "prompt", a name and your description; optionally a sample script of up to 2,000 characters. The response returns only an ID, a name, a date and your description: not one second of audio. Save the ID.
2. Open a GPT-Live session and set that ID as the voice.
3. Configure the delegation as Decisions, with your closed-choice questions.
4. Send your app's current state, wait for the session to confirm it and, when the state changes, delete the old state and send the new one.
5. Read the answer from Decisions by the delegation ID.
6. Run the action, verify it happened and tell Live.

The official example is a car: the driver says "make it warmer" and the cabin goes from 70 to 71 degrees Fahrenheit.

What a call costs

GPT-Live charges $0.05 per minute, billed by the second. A 3-minute call is $0.15; 100 calls a day, $15; 30 days, $450 a month. That is only the voice layer: the model that reasons behind it is billed separately and the changelog publishes no price for Decisions. If a call that used to be lost brings you a $20 sale, that sale pays for 133 three-minute calls (our assumption, not a data point). The fair comparison is not against a receptionist, but against the call nobody answers today.

Photo: Bjoertvedt, CC BY-SA 3.0, via Wikimedia Commons.

Three business cases (our analysis)

These cases are WAiBOT-AI's analysis, not cases published by OpenAI.

• A dental office with one receptionist. A patient calls at nine at night; the agent understands "I need to move my Thursday appointment" and Decisions picks reschedule while she is still explaining why.
• A restaurant taking phone orders at rush hour. For each dish, Decisions answers add, remove or nothing, and the app updates the total from the real order state, not from the conversation's memory.
• An agency with a different voice per client. One description per business, saving each voice's ID and description. The cap is 100 voices per organization, including those your application creates for your clients: with 30 clients you are fine, but client 101 does not fit. Voices are adult-only and you cannot ask them to imitate a named person.

Who this is not for

If you need your own cloned voice, that is a different path: it goes through OpenAI sales, a consent recording and a cap of 20 voices. If you already built on Realtime or Text to Speech, prompt-generated voices do not work there: they exist only inside GPT-Live. And if nobody on your team can program, there is a Playground for creating voices, but the flow with actions is through the API.

The objection: don't people prefer talking to a person?

Yes. A Kinsta survey reported by No Jitter says 93% of US consumers prefer a human to an AI (a high third-party figure: treat it with caution). But for a small business the real comparison is usually AI versus nobody answering. The objection wins on delicate calls, such as a complaint, an emergency or a large charge: there the agent should hand the call to a person. OpenAI also requires disclosing that the voice is AI-generated, and that disclosure belongs in the first second.

What you are not being told

OpenAI claims GPT-Live performs 30 percentage points better than its previous real-time model on a simultaneous-conversation benchmark. That is a vendor claim: to truly believe it, someone independent would need to repeat it with real calls, in Spanish and with noise. Also, this week's nine new voices are for German, French, Hindi, Japanese and Korean; none is declared as Spanish. What accent you get when you describe a voice in Spanish is not documented, and the API response includes no sample audio.

Where it breaks

• Every request saves a new voice: repeating the same description gives you a different voice and uses one of your 100 slots.
• The voice is created blind and you only hear it once you open a session.
• The guide recommends repeating the voice description in the session instructions.
• When Decisions fires from the transcript, results can arrive out of order: apply only the most recent one.

How to apply it in your business?

1. Save each voice's ID and description the same day you create it.
2. Repeat the description in the session instructions.
3. Design your Decisions questions with closed choices: add, remove or nothing.
4. Do the math with $0.05 per minute plus the reasoning model, and set a spending cap before opening it to the public.
5. Disclose from the first second that it is an AI voice, and leave a way out to a person.

Before selling it, make a test call in Spanish and with noise. For another voice-over-phone case, read the full breakdown of Grok Voice 3.0 and its price per minute.

Frequently asked questions

How much does it cost to create a voice in GPT-Live?

Creating the prompt-generated voice is free; speaking with it is billed at the regular GPT-Live rate of $0.05 per minute, plus the model that reasons behind it.

Can I hear the voice before using it?

The API response returns only metadata (ID, name, date and your description), with no sample audio.

How many prompt-generated voices can I have?

Up to 100 per organization, including those your application creates for its clients.

Does it work in Spanish?

The nine new voices are for other languages and the accent you get when describing a voice in Spanish is not documented: test with a real call.

Can I clone my own voice with it?

No. A cloned voice requires OpenAI sales, a consent recording and has a cap of 20 voices.

Conclusion

This week GPT-Live stopped being a voice demo and became something that can work in a business: a GPT-Live voice agent with a voice of its own that does things while you talk. The question that remains is which call your business loses today. Watch the full video above and tell us what first job you would give your voice agent.