OpenAI’s free prompt voices only speak inside GPT-Live

So your support bot needs a voice that isn't one of the stock presets everybody else ships. As of 16 October that's one API call: you POST a description to /v1/audio/voices, OpenAI hands back an id like voice_123, and creating it costs nothing. We read the custom voices guide, the GPT-Live session docs and the pricing page side by side. The feature's fine. It's where the voice can't go that'll catch teams out.

What shipped on 16 October

The changelog entry is short. Prompt-to-voice is generally available to all API customers, through the API or the voice Playground, for the gpt-live-1 model. The same day OpenAI added nine preset voices: Nuri, Juni and Flitz for German, Brise and Sillage for French, Nira for Hindi, Harema and Shida for Japanese, and Noeul for Korean. A day earlier, on 15 October, GPT-Live sessions gained native Decisions integration, so an app can trigger actions while the user's still talking.

The request itself is small. You send type: "prompt", a name, a prompt describing accent, pitch, tone and pace, and model: "auto". There's an optional script_hint of up to 2,000 characters of representative speech, which is how you tell it the voice will mostly read order numbers rather than bedtime stories. The guide doesn't say what "auto" picks. Each organization can hold up to 100 prompt-generated voices.

Voices built from a recording are a separate path with tighter rules: 20 per organization, samples of 30 seconds or less, a consent recording read word for word, and access limited to "eligible customers" through sales. That's the cloning product. Prompt voices skip all of it, no reference audio and no consent clip, which is exactly why they're open to everyone.

The catch is that the voice lives in one API

Here's the line we'd print out. The guide says prompt-generated voices work with GPT-Live and "aren't supported by the Text to Speech, Realtime, or Chat Completions APIs." Cloned voices work in all four. Prompt voices work in one.

Why does that matter? Plenty of voice stacks aren't a single API. A phone line might run live conversations on GPT-Live, then fall back to the Text to Speech endpoint for an outbound reminder call or a cheap batch of prerecorded prompts. With a prompt voice, those pieces can't share a voice. Your agent sounds like the brand persona you designed, and the reminder call sounds like somebody else. If brand consistency across channels is the point of a custom voice, that's a real gap, and the only documented fix is the recording path that most customers can't get into without talking to sales.

There's a second constraint in the session docs. The voice is set at session start, through session.audio.output.voice.id, and it can't change mid-session; you start a new session to switch. That's fine for one persona per call. It's awkward if you planned a handoff between two voices inside the same conversation.

And there's no preview. The creation response carries metadata only, it "doesn't include audio or a preview." To hear what you made, you open a GPT-Live session with it. GPT-Live bills $0.05 a minute, with backend model and tool tokens charged on top at their own rates. Auditioning is cheap (thirty seconds of a candidate voice is two and a half cents), so the money isn't the problem. The loop is: describe, create, open a session, listen, tweak the wording, create again. With a 100-voice ceiling per organization and no delete call shown in the guide we read, a team iterating carelessly could burn through that quota faster than it expects. We might be wrong about deletion, the guide may simply not cover it, but we'd check before handing the Playground to a marketing team.

What we'd check before shipping one

Compliance first. Voice creation through /v1/audio/voices isn't compatible with Zero Data Retention, and that sentence covers prompt voices as well as recordings. If your organization runs under ZDR, read that with whoever signed the contract before anyone creates a voice.

Then the usage rules, which are stricter than "be nice." You describe qualities, not people: no named person, public figure or fictional character. No children's voices. Listeners must be told the speech is AI-generated, and any provenance metadata or watermark has to stay intact. The guide also warns that voices "aren't guaranteed to be unique," so the warm mid-range American voice you described could land close to a competitor's warm mid-range American voice. Don't build a trademark around it.

Finally, keep your prompts. OpenAI's own advice is to save the returned id and retain the original description. We'd go further and store the full request, script_hint included, next to the id in version control. The docs don't promise that sending the same prompt twice gives the same voice, and we haven't tested it, so treat the id as the asset and the prompt as its documentation.

Our take: for a GPT-Live-only product, this is a good deal and an easy upgrade over a preset. For anything that mixes live conversation with batch speech, it's a trap that looks free. Pick the voice path by where audio gets generated, not by which demo sounds nicer. We list voice model prices alongside text models in our LLM API pricing tracker.

FAQ

Sources: OpenAI, API changelog (15 and 16 October 2026); OpenAI, Custom voices guide; OpenAI, Managing GPT-Live sessions; OpenAI, API pricing; OrcaRouter, GPT-Live-1 prompt-to-voice and nine new voices.