GPT-Live-1 Voices Are Now Something You Describe: Prompt-to-Voice Ships and Nine International Presets Land

For most of the year, choosing a voice for G​PT-Live-1 meant picking a string. You read a list, you set audio.output.voice/ to one of them at session start, and if none of them fit the brand you were building, you shipped with the closest one. That changed on 16 October 2026, when O​penAI released prompt-to-voice to all API customers and added nine natural voices aimed at five non-English markets. The new capability is small on paper — one endpoint, free to call — and it rewrites the earliest decision most voice products make. It also lands a day after G​PT-Live-1.1 picked up native Decisions integration inside the same Live sessions, which matters because the two changes push in opposite directions: one makes the voice more bespoke, the other makes the thing behind the voice more active.

Here is what actually shipped, what it costs, and the two asymmetries in the new voice list that nobody's launch note mentions.

What the October 16 release contains

Two things, both tagged to the gpt-live-1/ model and to two endpoints — v1/audio/voices/ and v1/live/sessions/.

• Prompt-to-voice — create a custom voice from a text description, through the API or the voice Playground, then reuse it in G​PT-Live sessions. Released for all API customers.

• Nine new natural voices — Nuri, Juni and Flitz (German); Brise and Sillage (French); Nira (Hindi); Harema and Shida (Japanese); Noeul (Korean).

• Cost — creating a prompt-generated voice is free. Using it speech-synthesises at regular G​PT-Live session rates, and auditioning it in the Playground counts as a normal session for billing.

• A day earlier, on 15 October — native Decisions integration in G​PT-Live, so an app can act while the user is still speaking or react to what Live says, via native delegation or transcript deltas.

The pricing context that makes those numbers legible: a G​PT-Live-1 voice session runs $0.05 per minute, billed by the second and not rounded up. That line is unchanged by this release, and no part of prompt-to-voice changes it.

The API call is one POST, and its shape tells you what the model is doing

You send JSON to /v1/audio/voices/ with type: "prompt"/, a name/, the prompt/ description, and "model": "auto"/. An optional script_hint/ of up to 2,000 characters supplies example speech to steer the generation; leave it out and the service writes the script itself, though blank or very short scripts are rejected.

The response contains no audio. It returns metadata only — an object type, an id/, the name/, a created_at/ in Unix seconds, the type/, and your original prompt. That is a deliberate design choice with a practical consequence: there is nothing to preview at creation time, so you audition it the way you would use it, which bills as a session. Every request saves a brand-new voice, so keep the ID and the description together — the description is the only record of what you asked for.

The controllable qualities are all expressed as ordinary language rather than as separate parameters: accent, age, presentation, pitch, timbre, texture, loudness, speech rate and distinctiveness. O​penAI's own examples include an accent phrased as "American English, with a strong accent" and a pace phrased as "Medium pace." There is no numeric prosody slider anywhere in the surface, which means the prompt is the API.

That is a real tradeoff. A slider is reproducible; a sentence is not. Two teams writing the same paragraph of description may not get the same voice, and the documentation says outright that generated voices are not guaranteed to be unique. Keep the prompt text under version control the same way you keep your system prompt.

The limits worth designing around

• 100 prompt-generated voices per organisation. The cap applies to your customers too, so if you are generating a voice per end user, 100 is the ceiling on your whole tenant base, not a per-account allowance.

• G​PT-Live only. Prompt-generated voices do not work with the Text to Speech, Realtime, or Chat Completions APIs. A voice you describe today is a voice you cannot reuse in a batch narration pipeline tomorrow.

• Not usable with Zero Data Retention. Voice creation through /v1/audio/voices/, including the recording path, is excluded.

• Adults only, and no imitation. Children's voices are not permitted, and the guidance is to describe qualities rather than to name a real person or character.

• Recording-based voices are a separate, tighter path — a maximum of 20 per organisation, samples of 30 seconds or less in mpeg, wav, ogg, aac, flac, webm or mp4, at least five seconds of real speech and fifteen transcribed tokens, uploads capped at 10 MiB, with a consent recording required.

The 100-voice cap and the G​PT-Live-only restriction are the two that shape architecture. If you are a platform generating a voice per customer, you have a hard ceiling and no fallback path into the cheaper speech APIs.

The full preset list, and where it is thin

The nine additions land on top of a library that O​penAI's Live voice guide now lists twenty-one deep, plus marin/ as the default. Read as a whole, the list says something about who O​penAI expects to be building:

• English — six regional colours, not one. Quartz and Ripple are Australian; Vesper is British; Willow and Stone are Irish; Gleam and Meridian are North American; Beacon is Filipino; Delta and Cinder are Southern U.S. Four of those are marked Generated rather than Natural.

• Portuguese — Bossa and Tempo, both Brazilian, both Natural.

• German — Nuri, Juni and Flitz. Two feminine, one masculine.

• French — Brise and Sillage. One feminine, one masculine.

• Japanese — Harema and Shida. One feminine, one masculine.

• Hindi — Nira. Feminine only.

• Korean — Noeul. Feminine only.

Two asymmetries fall out of that. First, German gets three voices while French, Japanese, Hindi and Korean get two, two, one and one — an allocation that tracks O​penAI's European market priorities more than it tracks speaker populations. Second, Hindi and Korean each get exactly one voice, and both are feminine. If your product needs a masculine Hindi or Korean voice from the preset library, there isn't one, and your options are a recording-based voice or a prompt-generated one. That is precisely the gap prompt-to-voice was released to fill, arriving the same day — which suggests the sequencing is not a coincidence.

The guide also flags something easy to misread: "regional influence" describes a voice's speaking style, not a hard constraint on which languages it can speak. Test candidate voices against the actual languages and pronunciations your app encounters rather than assuming a German-labelled voice is the right choice for German input.

The setting you cannot change mid-call. Voice is a startup setting. You assign audio.output.voice/ when you create the session, and to change it you start a new session. Forks inherit the source session's model, original instructions and input, so comparing two voices means creating two sessions from the same configuration rather than switching inside one. If you are A/B testing voices against real users, that means two session factories, not one session with a flag — and it means every comparison costs full session time on both arms.

Where the backend sits, and why that is the cheaper half of the decision

The voice is the half of a Live session that users hear. The other half is delegation: G​PT-Live listens and speaks while a backend model or agent does the reasoning and tool work, and backend usage is billed separately at normal rates. O​penAI's own cost guidance makes the arithmetic blunt — full active session time is billable including while the backend works, and using a faster backend model with parallel tool calls so the user finishes a minute earlier saves $0.05 in voice charges. The voice rate is fixed per minute; the backend is the dial.

That is where a router earns its place rather than being bolted on. Client delegation lets you point a G​PT-Live session at your own agent, harness or service — which means the delegated backend can be any model, and the useful version of "any model" is one key and one endpoint rather than a contract per provider. OrcaRouter fronts 200-plus models at provider list price with no markup, so a backend price cut reaches you the same day it lands, and automatic failover covers a provider having a bad afternoon without the user hearing it. The routing DSL lets you compose the reasoning step out of several models instead of betting the conversation on one, which is the shape you want for a delegated backend that has to answer inside a live conversation's latency budget.

One boundary to state plainly: OrcaRouter does not host G​PT-Live-1, and nothing in this release changes that. The router's role in a Live architecture is behind the delegation boundary, not in front of the voice.

What to do this week

If you already ship a Live product on a preset voice, the nine additions are a free pass to test whether your non-English users are actually being served — swap the voice on a fraction of sessions and listen, remembering that the swap costs you a new session per arm.

If you were blocked on brand fit, prompt-to-voice is now worth a day of experimentation, with one caveat that will bite later: treat the description text as a build artefact. Version it, keep the returned ID beside it, and assume you will need to regenerate — because with a 100-voice organisation cap and no preview at creation time, the voice you liked on Tuesday is a thing you will have to be able to recreate on purpose.

And if you are building the whole system rather than just the voice, the more consequential October change may be the one nobody wrote a blog post about. A day before the voices landed, G​PT-Live picked up native Decisions integration — the ability to act while the user is still mid-sentence. A voice that fits is a product decision. A backend that answers before the user finishes asking is an architecture one.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.