OpenAI’s 16 October 2026 API update lets all API customers create a voice from a text description, then use it in GPT-Live. A recording is not needed for this workflow. That is useful for a small spoken guide, but it does not add a new voice button to every ChatGPT account.
What is available, and where?
The current custom-voices guide offers the API and voice Playground. Prompt-generated voices work with GPT-Live; they are not supported by the separate Text to Speech, Realtime or Chat Completions APIs. Check the destination of your existing application before creating a voice. A ChatGPT subscription alone does not establish API access, billing or regional eligibility.
Creating the voice is free; generating speech uses regular GPT-Live session billing. Repeated listening tests can therefore incur usage charges. The guide currently allows up to 100 prompt-generated voices per organization, including voices created for customers. Reuse a saved voice rather than creating another for every session.
Give the voice a job
For a fictional library kiosk, try this description: “An adult voice speaking clear English at an unhurried pace, with a warm tone and distinct pauses between opening hours.” Describe qualities instead of asking for a celebrity’s voice. The provider requires adult voices and disclosure that the speech is AI-generated.
Keep the voice description separate from the words the kiosk should say. First try a short invented announcement with a weekday, a time and a room number. Listen for the exact details; a pleasant voice can still read a number unclearly.
Check before replacing a working voice
Save the voice ID and its description. Compare the same short script with your existing voice at the same playback volume. Check names, numbers, pauses and intelligibility on the device your listener will use. If these fail, revise one quality and compare again. A voice is not guaranteed to be unique, and a successful description does not prove suitability for every language.