OpenAI lets you describe a voice, then reuse it in GPT-Live

An interactive demo needs a calm adult guide, but you do not have a recorded speaker to clone. OpenAI’s October 16, 2026 API update adds prompt-to-voice: create a saved voice from a written description and reuse it in GPT-Live. Voice creation is free; speaking with it uses normal Live billing. Dated announcement.

Save the voice, then judge what listeners hear

The API or voice Playground creates the voice resource. Save the returned ID and original description; creating the resource returns metadata, not a recording. The guide limits an organization to 100 prompt-generated voices, so reusing an existing ID also avoids unnecessary duplicates. Custom-voice guide.

An illustrative brief might ask for an adult guide with a warm tone, moderate pace and clear English pronunciation. A useful listening check would include a product name, a price and a correction midway through a sentence. Judge whether people can understand the information and interrupt the guide comfortably. We have not created or listened to this example.

A distinctive description is not evidence of an exclusive brand voice. OpenAI says generated voices need not be unique, requires adult voices and AI disclosure, and prohibits impersonation. Describe qualities you need rather than naming a real performer to copy. Documented restrictions.

Check the API before planning the production workflow

Prompt-generated voices work with GPT-Live, not the separate Text to Speech, Realtime or Chat Completions APIs. Listening in the Playground is a billed Live session too. This update therefore does not establish a ready-made voiceover export feature in ChatGPT’s consumer voice picker. Supported surface and billing.

Our Live API explainer covers the surrounding application, and the voice API comparison helps identify which surface your project actually uses. Start with that choice before creating a collection of voices.