OpenAI GPT-Live Adds Prompt-to-Voice: Pricing and Setup Guide
On October 16, OpenAI announced in its API changelog that prompt-to-voice is available to all API customers. You describe a voice in text, then use it in OpenAI GPT-Live voice sessions. Creating the voice is free, and generating speech with it is billed as a standard GPT-Live session. The same day, OpenAI also added 9 natural voices for German, French, Hindi, Japanese and Korean.
This article reflects the situation as of October 2026. The facts come from OpenAI's official changelog; check the official docs for exact fields and limits.
What is GPT-Live, and what changed?
GPT-Live 1 became generally available in the API on September 10 at v1/live/sessions. It handles full-duplex voice conversations. Usage of the backend model and tools is billed separately.
Recent changes (dates are from the official changelog):
| Date | Change |
|---|---|
| Sep 10 | GPT-Live 1 generally available, v1/live/sessions |
| Oct 6 | Decisions API enters beta, v1/decisions, used with GPT-6 series models |
| Oct 15 | GPT-Live natively integrates Decisions, so it can call Decisions directly |
| Oct 16 | prompt-to-voice opens to all API customers, v1/audio/voices |
| Oct 16 | 9 new natural voices added |
The new voices are: German Nuri, Juni, Flitz; French Brise, Sillage; Hindi Nira; Japanese Harema, Shida; Korean Noeul. The changelog does not list any Chinese voices.
How do you use prompt-to-voice?
OpenAI says you can create a voice from a text description through the API or the Playground, then reuse it in GPT-Live sessions. A minimal flow along those lines:
- Try your description in the Playground first and listen to the result. Move to the API once you are happy with it, so you are not iterating inside code.
- Call
v1/audio/voiceswith the description text and save the voice identifier it returns. Request field names are defined in the official API docs. - Reference that voice when you create a
v1/live/sessionssession. The exact way to reference it is also in the official docs. - Store the voice identifier in your configuration. Don't recreate the voice for every session.
A good description covers perceived gender, age, speaking pace, mood and use case, for example "calm, slightly slow customer-service tone". This is a writing tip, not an official parameter.
How is it billed?
The official pricing has three parts:
- Voice sessions cost $0.05 per minute, billed by the second.
- Usage of the backend model and tools is billed separately and is not included in that price.
- Creating a prompt-to-voice voice is free. Generating speech with it is billed as a standard GPT-Live session.
The changelog gives no pricing for the Decisions API or the Decisions integration, so confirm on the official pricing page before launch.
Here is an example (hypothetical numbers, not measured): a customer-service line takes 200 calls a day, averaging 3 minutes each.
- Voice session portion: 200 × 3 × $0.05 = $30 per day.
- Backend model and tool costs are added on top, depending on the model you choose and how often it is called.
So estimate voice minutes first, then estimate the backend portion separately. Don't stop at $0.05.
How sensitive is realtime voice to network latency?
In a full-duplex conversation, round-trip time becomes "how long until the other side responds". When you reach OpenAI's API from mainland China, fluctuations in the network path are easier to hear than they are in ordinary text requests. To check:
- Open the cdnping homepage speed test and look at the "first connection" and "round trip" values for api.openai.com. Round trip is the median of 3 requests over a reused connection, which is closer to what a realtime session feels like.
- Test several times at different hours. A single result tells you little.
- Check the OpenAI status page for the last 30 days of probe records on the official endpoint, to rule out server-side problems.
- For the causes of latency and how to judge it, see Latency when reaching AI APIs from China.
Note that OpenAI services are only available in supported regions. Confirm that your region is supported before use, and comply with the terms of service and local regulations.
What else to check before launch
- Access method: going direct to the official API, using a cloud provider's hosted service, and using an aggregation gateway differ a lot in billing and compliance. See How to choose an AI API access method.
- Bill reconciliation: multi-part billing (voice minutes plus a backend model) is easy to get out of sync. See Why your API bill looks wrong.
- Related pricing changes: if the backend model is from the GPT-6 series, see GPT-6 Sol / Luna API prices and cache billing.
- Agent-side updates: OpenAI Agents API adds Computer Use and subagent delegation.
Sources
- OpenAI API Changelog (GPT-Live, prompt-to-voice, new voices and Decisions API entries, consulted October 2026)
