Does OpenAI store audio sent to the Whisper API?

By default, OpenAI does not use data you send to its API to train its models ¹. Storage is a separate question. By default, OpenAI keeps abuse monitoring logs for up to 30 days ², and its own endpoint table says the transcription endpoint keeps none ³. The two statements conflict, and OpenAI's data page does not settle which one covers the transcription endpoint. Until OpenAI confirms the answer for your account in writing, treat the audio as kept for up to 30 days.

OpenAI does not train on API data unless you opt in

OpenAI's data controls page says that since March 1, 2023, data sent to the API has not been used to train or improve its models, unless you explicitly opt in to share it ¹. That answers the training question only. The same page lists two separate kinds of stored data, abuse monitoring logs and application state, so a question about training does not answer a question about storage.

Abuse monitoring logs can be kept for up to 30 days by default

Abuse monitoring logs are generated for API feature usage. By default they are kept for up to 30 days, unless the law requires longer or OpenAI needs them to protect its services ². The page does not say whether an audio file is included in those logs, so do not assume that it is left out.

The transcription endpoint's table says no retention, which conflicts with the general rule

OpenAI's per-endpoint table lists /v1/audio/transcriptions with no data used for training, no abuse monitoring retention, no application state retention, and eligibility for Zero Data Retention ³. The table is the more specific statement on the page, and it covers the transcription endpoint directly. The general rule above the table says that all API feature usage generates logs kept for up to 30 days.

The page does not reconcile the two. It also names endpoints, not models, so the table does not confirm how the whisper-1 model is handled. Read the row as OpenAI's stated behaviour for that endpoint, and ask OpenAI in writing before you rely on it for records that carry legal or medical weight.

Zero Data Retention needs OpenAI's approval first

The controls that exclude your content from abuse monitoring are called Zero Data Retention and Modified Abuse Monitoring. OpenAI says they are currently subject to prior approval and to acceptance of additional requirements ⁴. Without that approval, you cannot switch them on from a settings page. Until you have that approval, assume the 30-day default.

Apps that use your own key follow your agreement with OpenAI

Some Whisper apps let you enter your own OpenAI key, and then the app is only a client of your account. Superwhisper states this for users who supply their own keys. It acts as a client of the provider account, and your agreement with the provider governs the data ⁵. So the OpenAI terms and your account settings matter more than the app's privacy page. Our question on whether Whisper apps send audio to OpenAI explains how to tell which path an app takes.

What to do: check your account, then get the answer in writing

1. Decide whether the recordings can go to OpenAI at all. For regulated material, read whether Whisper AI is HIPAA compliant first.
2. Review your organization's data controls in the OpenAI dashboard, and write down the settings you see and the date.
3. Ask OpenAI whether the transcription endpoint and the model you use keep no abuse monitoring logs for your organization.
4. Keep the reply. If the documentation and your account screens disagree, the written answer is the record you can rely on.
5. Delete the audio and transcripts you no longer need. The guide to voice data retention compares what the other large vendors keep.

For the full picture of the Whisper model and its privacy options, read whether Whisper AI is safe. The overview of speech to text privacy covers the same questions across vendors.