Is OpenAI shutting down the Whisper API?

Yes, the hosted Whisper API is being retired. On August 26, 2026, OpenAI notified developers that whisper-1 would be removed from its API, and it set the removal date at February 26, 2027 ¹. After that date, requests that name whisper-1 will stop working. If your code, app or automation calls whisper-1, you need to switch models before then.

Two things are not shutting down. The open-source Whisper model is a separate download that you run yourself, and its code and weights carry the MIT License ². And OpenAI's transcription service is not ending: it is moving to newer models. This page covers the API change. Our guide to what Whisper AI is explains the model itself.

The notice covers four models, not only whisper-1

OpenAI's entry on its deprecations page is titled "Transcription models". It lists four models for removal on the same day: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize ³. So the question "is the Whisper API shut down?" has a wider answer than the name suggests. The newer gpt-4o models that many developers moved to after whisper-1 are on the same schedule.

If you searched for "gpt-4o-transcribe deprecation", the answer is the same: it is deprecated with whisper-1, with the same February 26, 2027 date.

OpenAI's recommended replacements are gpt-transcribe and gpt-live-transcribe

OpenAI's table lists gpt-live-transcribe or gpt-transcribe as the recommended replacement for whisper-1, and for the other three models too ⁴. Its speech-to-text guide now names gpt-transcribe first and calls it the recommended model for transcribing recorded speech in its original language ⁵. Its quickstart sends audio to the same transcriptions endpoint with the model name gpt-transcribe ⁶.

In practice, for recorded files the migration is often a change of model name. Live, streaming transcription is priced separately: gpt-live-transcribe is listed at $0.017 per minute ⁷.

Price goes down for recorded audio

whisper-1 costs $0.006 per minute ⁸. gpt-transcribe is listed at $0.0045 per minute ⁹, which is cheaper for the same recording. Our question on OpenAI's speech-to-text API pricing compares the models in more detail.

Accuracy is a reason to test, not to worry

On Artificial Analysis's leaderboard, gpt-transcribe had a word error rate of 3.3 percent, against 4.0 percent for gpt-4o-transcribe ^{10 11}. A benchmark average does not predict your accents, microphones or vocabulary, so run a sample of your own audio before you switch.

The translation endpoint is also going away

If you use whisper-1 to write English text from foreign-language speech, note a second change. The same table lists the /v1/audio/translations endpoint for removal on February 26, 2027, with v1/realtime/translations as the recommended replacement ¹². This is not a model-name swap: it is a different endpoint, so the code needs a real change. See our question on whether Whisper can translate.

What you should do depends on how you use Whisper

The right move depends on where Whisper runs in your setup.

• You call the OpenAI API with whisper-1. Change the model to gpt-transcribe (or gpt-live-transcribe for live audio), test it on real recordings, and finish before February 26, 2027. Do not wait for the last week: a model that is removed returns errors, not a quiet fallback.
• You use an app built on Whisper. Check whether it sends audio to OpenAI or runs the model on your computer. Apps that run it locally are not touched by this change. Apps that call the API need their maker to update. Our question on whether Whisper apps send audio to OpenAI explains how to tell.
• You run the open-source model yourself. Nothing changes. It needs no API key, and the question on running Whisper offline covers the setup. Keep in mind that the weights you downloaded do not expire.
• You want to avoid depending on one vendor's API. Run a model on your own machine. On-device speech recognition has no vendor shutdown date. Yap, which we make, is dictation software that works this way on Mac and Windows: see the download page. If you only need occasional transcripts, free file transcription in the browser may be enough.

Why a deprecation is not a surprise

Hosted models are versions of a service, not products you own. A provider can retire a model while the underlying technology continues, as this notice shows: whisper-1 is retired, and the open-source Whisper weights remain under their license. The lesson for any project is to pin the model name in one place in your code, so a change like this is a one-line edit.

Dates can move, so check the source

This page reflects OpenAI's deprecations page as read on October 27, 2026. Deprecation dates are set by the vendor and can be revised. Read OpenAI's deprecations page before you plan, and see the question on the best Whisper alternatives if you want to compare providers. For the model's broader background, return to the speech recognition overview or the question on whether Whisper has an API.