Do you need an OpenAI API key to use Whisper offline?

No. The open-source Whisper model runs offline without an OpenAI API key, because your own computer does the recognition. The README for the openai-whisper package describes transcription as a command you run on an audio file, and it mentions no key. An API key matters only when you use OpenAI's hosted transcription service, which runs on OpenAI's servers and cannot work offline. The apps to check are the ones that ask for a key, since a key usually means your audio goes to OpenAI.

The open-source model needs no key, because you run it yourself

The openai-whisper README introduces command-line use with the line "The following command will transcribe speech in audio files" ¹. The example names audio files and a model, and it contains no sign-in step and no API key. The first time you load a model by name, the package fetches the model file with urllib, a standard library call, unless the file is already in its cache folder ². That download is the only network step, and it fetches the model, not your audio.

Once the model is on your computer, the package transcribes from the file. You do not have to create an account at OpenAI to run it. That is why the offline answer does not depend on a key. For the steps to set up an offline run, see our answer on whether Whisper AI can run offline.

The key belongs to OpenAI's hosted API, which sends your audio away

OpenAI's hosted transcription API requires your secret key on every request. The key goes in an Authorization header, written as the word Bearer followed by the key ³. Each request uploads the audio to OpenAI, so the key and your recording travel together. Hosted transcription is also billed by the minute. OpenAI's price list, as we checked it on 16 October 2026, shows its gpt-transcribe model at $0.0045 per minute ⁴. That is OpenAI's price for its own hosted model, not a charge for the open-source package, which has no per-minute fee.

A hosted API cannot run offline, because it needs a connection to OpenAI for every request. If you need a key for your work, read our question on whether OpenAI stores Whisper audio before you send any recording.

Apps add keys and servers, so test an offline claim yourself

Apps built on Whisper make their own choices, and the key question gets confusing there. Whisper To Input, an Android keyboard, does its recognition on a backend that you select. The two options are the OpenAI API with your own key, or a Whisper server that you host yourself ⁵. The first needs a key and a connection. The second needs no OpenAI key, but the phone still has to reach your server, so the phone is not offline.

Some apps need no key and still work offline. Handy describes itself as a free, open source speech-to-text app that works completely offline ⁶. whisper.cpp shows Whisper running fully offline on a device, in its iPhone 13 demonstration ⁷.

The absence of a key is not proof of privacy. An app can send audio through its own servers and need no key from you. Superwhisper states that with cloud transcription, audio goes to its proxy, is transcribed and comes back ⁸. So check the processing mode as well as the key field.

What to do if an app asks for an OpenAI key

1. Find the key field, and decide whether you want it. If the app asks for an OpenAI key, it can send audio to OpenAI under your account.
2. Look for a local or on-device mode in the settings, and choose it for sensitive recordings.
3. Download the model while you are online. Then switch to airplane mode and transcribe a real recording. If the text still appears, the work happened on the device.
4. Read the privacy terms for any cloud mode, and check the retention rules of the provider whose key you enter.

For the wider privacy picture, read the guide to whether Whisper AI is safe. Two related questions are whether Whisper apps send audio to OpenAI and whether the Whisper model is safe to download and run. The overview of speech to text privacy explains where a key sends your words, and the glossary defines bring your own key and on-device speech recognition.