OpenAI Releases Prompt-to-Voice for All API Customers, Plus Nine GPT-Live Voices

OpenAI’s API changelog for 16 October 2026 says the company has released prompt-to-voice for all API customers and has added nine natural voices for GPT-Live. The matching Custom voices guide marks prompt-to-voice as generally available to all API customers.

TSN found no OpenAI blog post and no OpenAI X post announcing this drop. The confirmed source set is the changelog and developer guide only. Aggregator write-ups are not used here.

This is a separate developer feature from ChatGPT’s GPT-Live-1.1 voice rollout. TSN’s earlier GPT-Live-1.1 / ChatGPT Voice post recorded that API access was then described as “coming soon.” The 16 October changelog is the docs-side confirmation that prompt-generated custom voices are now creatable and reusable in GPT-Live sessions for API customers.

What the changelog says

Under Oct 16, OpenAI writes:

Released prompt-to-voice for all API customers. Create a custom voice from a text description through the API or voice Playground, then reuse it in GPT-Live sessions. Creating a prompt-generated voice is free; speech generation uses regular GPT-Live session billing.

And:

Added nine natural voices for GPT-Live: Nuri, Juni, and Flitz (German); Brise and Sillage (French); Nira (Hindi); Harema and Shida (Japanese); and Noeul (Korean).

Those nine names also appear on OpenAI’s live-conversations voice options docs. TSN makes no quality claims about how they sound.

For billing context already on the changelog (Sep 10): GPT-Live 1 voice sessions “cost $0.05 per minute, billed per second”; backend model and tool usage is charged separately. Prompt-voice creation is free; using the voice in a session is ordinary GPT-Live session billing.

How prompt-to-voice works (guide)

The guide says you create a voice with POST /v1/audio/voices and "type": "prompt", or through the voice Playground. You describe vocal qualities in plain language (accent, age presentation, pitch, timbre, pace, and so on). An optional script_hint can supply “up to 2,000 characters” of representative speech; if omitted, the service generates a script from the description.

Each successful create saves a new voice (metadata and ID; the create response does not include audio). You then set session.audio.output.voice.id in a GPT-Live session and, for performance guidance, include the original description in session.instructions.

Hard limits and responsible-use rules

Quote the constraints as OpenAI states them:

• GPT-Live only. “Prompt-generated voices work with GPT-Live. They aren’t supported by the Text to Speech, Realtime, or Chat Completions APIs.”
• Cap: “You can create up to 100 prompt-generated voices per organization.” The limit also applies when an application creates voices for its customers.
• No children’s voices. “Children’s voices aren’t allowed”; describe an adult voice.
• No named-person imitation. “Don’t request imitation of a named person, public figure, or fictional character.”
• Uniqueness. “Voices aren’t guaranteed to be unique. Others may generate similar voices or output.”

The guide also tells developers to use voices responsibly, stay transparent that speech is AI-generated, and preserve provenance metadata or watermarks where present. Those are policy lines from OpenAI, not TSN assessments of enforcement.

Not the same as recording-based custom voices

OpenAI’s guide still documents a separate path: create a voice from a recording (consent recording plus reference sample). That route is limited to eligible customers, has its own caps, and — unlike prompt-to-voice — can be used with GPT-Live, Text to Speech, Realtime, or Chat Completions audio. This post is about the prompt-to-voice GA and the nine named GPT-Live voices, not about expanding the recording programme.

Why it matters

For builders of third-party voice agents, the gap called out in 22413 was API access after ChatGPT got new Live voices. The changelog now documents text-described custom voices that stick to GPT-Live session billing, plus a multilingual batch of stock voices. Related speech-stack context includes SpaceXSI’s Grok Voice 3.0 and OpenAI’s separate GPT-6 Pro ChatGPT upgrade.

What’s still unknown

• Whether OpenAI will publish a blog or social announcement (none found at draft time)
• How closely prompt-generated voices match a given description across accents and long sessions (no TSN listening tests)
• Whether the nine new stock voices will later appear outside GPT-Live

What this means

OpenAI’s API docs, not a blog or X thread, say prompt-to-voice is GA for all API customers, creation is free, speech rides GPT-Live billing, use is GPT-Live-only, organisations get at most 100 prompt voices, and children’s voices / named-person imitation are out of bounds — with uniqueness not guaranteed. Nine new GPT-Live stock voices (German, French, Hindi, Japanese, Korean) ship in the same Oct 16 changelog entry.

Sources

• [1] OpenAI API Changelog — Oct 16, 2026
• [2] OpenAI API Docs: Custom voices
• [3] OpenAI API Docs: Live conversations — voice options

Related stories

• OpenAI Rolls Out GPT-Live-1.1 and New Voices in ChatGPT Voice
• SpaceXSI Releases Grok Voice 3.0, Leading Artificial Analysis’s Speech-to-Speech Index
• OpenAI Upgrades GPT-6 Pro in ChatGPT with Parallel Agents