OpenAI custom voices: creation, session costs and retention
OpenAI’s October 16, 2026 API update made text-described custom voices available to all API customers. The practical distinction is between saving a voice and paying for conversations that use it.
Choose the right creation route
The custom voice guide distinguishes two workflows:
| Route | Required input and access | Where it works |
|---|---|---|
| Text description | All API customers. No reference or consent recording. | GPT Live only |
| Recording | Eligible customers, consent and reference recordings | GPT Live, Text to Speech, Realtime and audio Chat Completions |
Prompt-based creation is free, but Playground listening starts a billable session. Each successful request saves a new voice, returning metadata rather than audio. Keep its identifier and prompt. The organization limit is 100 prompt-generated voices, including customer-created ones. OpenAI requires adult voices and synthetic-speech disclosure, prohibits impersonation and does not guarantee uniqueness.
Budget for the conversation, not just the voice
The pricing page checked November 2 lists GPT Live 1 at USD $0.05 per minute. Sessions are billed per second, without rounding up to whole minutes. Backend model and tool usage is charged separately.
For example, ten minutes at that listed voice-session rate is $0.50 before backend charges. This is arithmetic, not a measured application bill. A cost comparison should record session duration and delegated work separately rather than treating free voice creation as free operation.
Select a voice before starting
The session guide sets voice selection at startup. Changing it requires a new session. Its nine additional natural presets cover German, French, Hindi, Japanese and Korean. Regional descriptions describe speaking style, not a substitute for pronunciation evaluation.
Suggested checks: names, quantities, technical terms and disclosure comprehension. No voices were tested.
Connect the voice to outbound phone calls
Phone deployment guidance updated November 9, 2026. OpenAI’s November 2 API update opened outbound SIP calling to all organizations. A developer can connect GPT Live to a phone number through a separate telephony provider.
The SIP guide requires a trunk supporting TLS signaling and SDES-SRTP audio. Destinations must be E.164 phone numbers, not SIP URIs. Create the call through the Live API session endpoint and choose the voice and delegation mode at creation. Your provider handles the phone network connection.
A successful creation response only confirms session initialization. Attach the sideband WebSocket connection immediately to monitor ringing, answered and failed events. Its replay window covers only the preceding three seconds. Deduplicate repeated events by their event identifier so a business action is not performed twice.
OpenAI limits ringing to three minutes and connected calls to two hours. Keep the backend connected until session closure returns final usage. Those limits and the provider’s own charges belong in your deployment plan alongside the voice-session price above.
Before using customer numbers, test with numbers you control. Include a busy line, an unanswered call and a failed connection. Check that the application records the outcome before deciding whether to retry. For an appointment reminder, successful dialing alone should never mark the appointment as confirmed. This is a suggested test plan. We have not tested outbound calls.
Voice storage and session storage differ
OpenAI’s data controls table lists voice creation as ineligible for Zero Data Retention, with application state kept until deletion. GPT Live sessions are eligible subject to limitations. Check both stages rather than assuming a conversation setting covers the saved voice.
Session storage is off by default. Where a project enables storage, saving a completed session retains its recording and index for 30 days. Zero Data Retention forces storage off and prevents recording downloads and forking. Delegated models and tools have separate controls. These documented distinctions are not a blanket privacy or compliance guarantee.
This API workflow is separate from the consumer experience covered in ByteForward’s ChatGPT Voice guide. Choosing a sound in ChatGPT does not describe how a developer creates and manages an API voice.
AI-generated conceptual illustration with the official OpenAI wordmark composited by ByteForward. OpenAI owns its mark. The sculptures are not actual products or audio measurements.
