GPT-Live Custom Voice vs Preset Voices: Which to Pick and Why

Pick a preset natural voice when you need to ship today, want a voice tuned for a specific language, and don't need a brand-specific sound. Pick a prompt-generated custom voice when the voice itself is part of the product and nothing built in fits. As of October 16, 2026, OpenAI's API changelog says creating a custom voice is free and speech from it bills at the regular GPT-Live session rate, so price alone won't decide this for you.

The real difference is one extra setup step. Natural voices are ready the moment you open a GPT-Live session. A custom voice needs a text description and a creation step first, either through the v1/audio/voices endpoint or the voice Playground. After that you reuse it in GPT-Live sessions like any other voice.

GPT-Live custom voice vs preset voices at a glance#

Everything below expands on this table. If you only wanted the decision, the last section is a checklist you can paste into a ticket.

What a natural voice is in GPT-Live#

GPT-Live 1 is OpenAI's full-duplex voice model. It went generally available in the API on September 10, 2026, running on gpt-live-1 through v1/live/sessions. The model handles the conversation while a backend model or agent does reasoning and tool calls, either through Responses delegation with an OpenAI model or client delegation to your own backend.

Natural voices are the built-in options you select for that session. On October 16, 2026, the changelog listed nine new ones, grouped by language:

• German: Nuri, Juni, Flitz
• French: Brise, Sillage
• Hindi: Nira
• Japanese: Harema, Shida
• Korean: Noeul

The point of a preset is that somebody else already did the tuning. You get a voice that sounds right for the language it was built for, with no creation step and nothing extra to manage. The trade-off is that you share it with every other developer who picked the same name. If you want the full list and a selection walkthrough, the companion piece on GPT-Live voices by language and how to select one covers it.

What a prompt-generated custom voice is#

Prompt-to-voice, released to all API customers on October 16, 2026, lets you describe a voice in text and get a voice back. You create it once, through the API or the voice Playground, then reference it in GPT-Live sessions. The changelog lists the endpoints involved as v1/audio/voices for creation and v1/live/sessions for use.

Two things stand out in how OpenAI describes it. First, the input is a text description. The changelog doesn't mention audio samples or cloning, so plan on writing a description, not uploading a recording. Second, the voice is a reusable asset. You don't regenerate it per session; you make it and then point sessions at it.

That makes a custom voice the right tool when the voice is a product decision. A branded assistant, a character in a game, a tone that's deliberately unlike the defaults: these are the cases where a preset will always feel like a compromise. The step-by-step creation flow lives in the prompt-to-voice API walkthrough, so this article stays on the comparison.

Setting up a preset voice vs a custom voice#

Here's the honest comparison of effort. Neither path is long, but one has a step the other doesn't.

Preset natural voice:

1. Decide which language your users speak.
2. Pick a voice name from the list for that language.
3. Select that voice when you create the GPT-Live session.
4. Test a short conversation and adjust your prompt, not the voice.

Prompt-generated custom voice:

1. Write a text description of the voice you want. Be concrete about tone, pace and character rather than vague adjectives.
2. Create the voice through v1/audio/voices or in the voice Playground. Creation is free, per the changelog.
3. Listen to it. If it's wrong, rewrite the description and create again; you haven't spent anything yet.
4. Reference the finished voice in your GPT-Live sessions.
5. Test the same short conversation you'd test with a preset.

Step 3 is where I'd spend the time. Because creation doesn't bill, iterating on the description costs you minutes, not money. Speech only starts billing once a session runs, which is the same moment a preset would start billing.

One practical tip: do the first few iterations in the voice Playground, where you can hear results immediately, and only move to the API call once you know what description works. The Playground and the API produce the same kind of reusable voice, so nothing is lost by starting in the UI.

Language coverage: which option fits your users#

This is where the two options differ most clearly in what's documented.

Natural voices are listed by language. The October 16, 2026 batch adds voices for German, French, Hindi, Japanese and Korean, and the changelog names each one. If your users speak one of those languages, you have a known-good starting point with a name you can put straight into a session.

Prompt-generated voices have no language table in the changelog. OpenAI's announcement describes creating a voice from a text description and reusing it in GPT-Live, but doesn't say which languages a generated voice speaks well or whether the description can target a language. Until the official docs list that, treat it as something to verify yourself: create the voice, run a session in your target language, and listen.

My recommendation follows from that gap. If a language-specific preset exists for your audience, start there even if you eventually want a custom voice. You get a baseline that's known to work, and you can compare your custom result against it. If no preset is listed for your language, a custom voice is the only documented route to something closer, but budget time for listening tests.

// Natural voices GPT-Live added on October 16, 2026, grouped by language.
const naturalVoices = {
  German: ["Nuri", "Juni", "Flitz"],
  French: ["Brise", "Sillage"],
  Hindi: ["Nira"],
  Japanese: ["Harema", "Shida"],
  Korean: ["Noeul"],
};

function recommend(language, needsBrandVoice) {
  const presets = naturalVoices[language] || [];
  if (needsBrandVoice) {
    return `${language}: create a prompt-generated voice (brand voice requested)`;
  }
  if (presets.length > 0) {
    return `${language}: start with a natural voice (${presets.join(", ")})`;
  }
  return `${language}: not in the October 16 batch; check the full voice list, then consider a prompt-generated voice`;
}

const requests = [
  ["German", false],
  ["Hindi", false],
  ["Portuguese", false],
  ["French", true],
];

for (const [language, brand] of requests) {
  console.log(recommend(language, brand));
}

The script only knows the nine voices from the October 16 changelog entry. Older presets exist too, so a miss here means "check the full list," not "no voice available."

Billing: what each option costs#

The billing story is short, which is unusual for a comparison like this.

GPT-Live voice sessions cost $0.05 per minute, billed per second, according to the September 10, 2026 changelog entry. Backend model and tool usage is charged separately. That's the session rate regardless of which voice is speaking.

For custom voices, the October 16 entry adds two facts: creating a prompt-generated voice is free, and speech generation uses regular GPT-Live session billing. So a custom voice costs the same per second of conversation as a preset, and the creation step adds nothing. The changelog doesn't list any per-voice charge for natural voices either.

What this means in practice:

• Iterating on a voice description is free. Create, listen, rewrite, repeat.
• Your real cost is session minutes, driven by how long conversations run, not which voice you chose.
• The backend model behind the session bills on its own terms. If you delegate to an OpenAI model through Responses, that model's token pricing applies on top of the session rate. Check the OpenAI pricing page for the model you delegate to.

Per-second billing matters more than it sounds. A 90-second support interaction is billed as 90 seconds, not rounded up to two minutes. Here's a small calculator for the session portion only.

# GPT-Live session cost: $0.05 per minute, billed per second (changelog, Sep 10, 2026).
# Backend model and tool usage bill separately and are not included here.
RATE_PER_MINUTE = 0.05
RATE_PER_SECOND = RATE_PER_MINUTE / 60

def session_cost(seconds: int) -> float:
    return seconds * RATE_PER_SECOND

scenarios = {
    "90-second support call": 90,
    "5-minute product demo": 300,
    "1,000 calls x 4 minutes": 1000 * 240,
}

print("Session cost only; preset or custom voice makes no difference here.\n")
for label, secs in scenarios.items():
    print(f"{label:26} {secs:>8} s  ${session_cost(secs):,.4f}")

print("\nCreating a prompt-generated voice adds nothing to these totals.")

If your numbers look wrong, the usual reason is forgetting the backend model. The session rate is the floor, not the whole bill.

When I'd pick each one#

Start with a preset when:

• You're validating whether voice works for your product at all.
• Your users speak a language with a listed natural voice.
• The voice isn't part of your brand and nobody will notice it's shared.
• You want the fewest moving parts in the first release.

Move to a custom voice when:

• The voice is a brand or character decision that a stakeholder will review.
• You've tried the presets and none has the tone you need.
• You need a consistent voice across several products and want one asset to reference.
• Your target language has no listed preset and you're willing to run listening tests.

A sensible sequence for most teams: launch on a preset, keep the session code voice-agnostic, and swap in a custom voice later once you know people actually use the feature. Since a custom voice is just a reference you pass into the same session, the swap is a config change rather than a rewrite. If your session also reacts mid-conversation, the GPT-Live Decisions integration is independent of the voice choice, so you can design that part in parallel.

Common mistakes and gotchas#

Confusing the two "custom voices" features. The December 15, 2025 changelog entry mentions custom voice support for eligible customers on older audio snapshots such as gpt-4o-mini-tts. That's a different, gated feature. Prompt-to-voice for GPT-Live is the October 16, 2026 release, and the changelog says it's available to all API customers. Don't assume you need special access because you read the older entry.

Assuming a custom voice is a premium tier. It isn't, at least on what's documented. Creation is free and speech bills at the same session rate. If someone on your team is blocking a custom voice on cost grounds, the changelog is the citation to send them.

Expecting audio-sample cloning. The documented input is a text description. If your plan was to upload a recording of a voice actor, the feature as described doesn't do that.

Picking a custom voice when a language preset already exists. For German, French, Hindi, Japanese or Korean, there's a named preset you can test in minutes. Try it first. A custom voice with no documented language coverage is a riskier starting point than a preset built for that language.

Forgetting the backend model in cost estimates. The session rate covers the voice conversation. The model doing reasoning behind it, and any tools it calls, bill separately. Voice choice doesn't change this, but it's the part people miss when the invoice arrives.

Iterating on the voice in production sessions. Because creation is free and sessions aren't, do your voice iteration in the voice Playground or with creation calls, not by spinning up live sessions to hear each variant.

Hard-coding the voice deep in your session setup. Keep the voice as configuration so moving from a preset to a custom voice, or between presets, doesn't touch application logic.

Quick reference checklist#

• Language has a listed natural voice? Start with it. Nine new ones as of October 16, 2026: German (Nuri, Juni, Flitz), French (Brise, Sillage), Hindi (Nira), Japanese (Harema, Shida), Korean (Noeul).
• Voice is a brand or character decision? Create a prompt-generated voice from a text description via v1/audio/voices or the voice Playground.
• Creation cost: free, per the changelog.
• Speech cost: $0.05 per minute, billed per second, for any voice. Backend model and tools bill separately.
• Language coverage for custom voices: not listed in the changelog yet. Run your own listening tests.
• Iterate on descriptions in the Playground, then move to the API.
• Keep the voice as config so you can swap later without code changes.
• Access: prompt-to-voice is available to all API customers as of October 16, 2026, per the OpenAI changelog.

Quick answers