OpenAI Prompt-to-Voice: API, Cost and Limits

On October 16, 2026, OpenAI made prompt-to-voice generally available to API customers. Developers can describe a voice in plain English, create a reusable voice ID and select that voice when starting GPT-Live conversations. OpenAI also introduced nine named GPT-Live voices across German, French, Hindi, Japanese and Korean. This update is most relevant to teams building conversational assistants that need a consistent spoken identity across sessions.

API compatibility comes first

The key distinction is compatibility. Voices created from a text description currently work with GPT-Live, not the separate Text to Speech, Realtime or audio Chat Completions APIs. OpenAI also documents a recording-based custom-voice workflow with different eligibility and API support. Before building a voice interface, teams should choose their target speech API first rather than assuming every voice type works everywhere.

Create and reuse a custom voice

Developers can create a voice in the voice Playground or call /v1/audio/voices with type set to prompt, a descriptive name and a text description. The prompt can specify an adult voice's accent, pitch, timbre, texture and speaking pace. An optional script_hint supplies representative text up to 2,000 characters. The creation response returns metadata and a saved voice ID, not a playable audio preview. Store that ID for later sessions.

To use the saved voice, provide its ID in session.audio.output.voice.id when starting GPT-Live. OpenAI recommends keeping the original description in session instructions to help preserve the intended delivery. Voice selection is a session-start setting; switching to another voice requires a new session. Test the chosen voice with the same dialogue on your intended transport, and measure pronunciation, pauses, turn-taking and interruption handling.

Costs, quotas and built-in alternatives

Creating a prompt-generated voice is free, but playing it in the voice Playground and generating speech in GPT-Live incur regular session charges. OpenAI's documentation lists a limit of 100 prompt-generated voices per organization, including those created for customers. For applications with multiple personas, reuse saved IDs and measure total conversation cost instead of treating voice creation as the main expense.

The nine built-in voices introduced in the release are Nuri, Juni and Flitz for German; Brise and Sillage for French; Nira for Hindi; Harema and Shida for Japanese; and Noeul for Korean. Built-in options are worth testing before generating a custom voice: a generated voice is not necessarily clearer, more natural or better suited to a target audience. Compare alternatives on identical scripts and real multi-turn conversations.

Data handling and evaluation

Operational details matter. OpenAI says creation through /v1/audio/voices is not compatible with Zero Data Retention, although GPT-Live speech generation can be eligible when ZDR is enabled for a project. The documentation also requires transparency that speech is AI-generated and notes that generated voices are not guaranteed to be unique. Teams should review applicable terms, product disclosure and data-retention needs before a public launch.

A practical evaluation starts with one customer journey and two or three descriptions that differ in only one vocal quality. Create each voice once, save the IDs and compare them against a suitable built-in voice. Record task completion, user comprehension, language-specific pronunciation, session latency and spend. Select a voice based on observed results, not marketing claims about a particular accent or synthetic voice.