whisper-1 deprecation: move to gpt-transcribe by 26 Feb 2027
The whisper-1 deprecation lands on 26 February 2027. What replaces it, how transcription prices change, and the features with no named replacement.
The whisper-1 deprecation is dated: OpenAI will remove whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize from the API on 26 February 2027, 158 days after 21 September 2026. The named replacements are gpt-transcribe for recorded files and gpt-live-transcribe for live audio. Plain transcription moves easily and gets cheaper. Word timestamps, subtitles, translation and speaker labels are harder, because OpenAI's own guide still points those at the models being retired. This post covers the mapping, the price changes, and how to test accuracy on your audio before you switch.
| Key fact | Detail |
|---|---|
| Announced | 26 August 2026 |
| Shutdown | 26 February 2027 |
| Models affected | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize |
| Replacement for recorded files | gpt-transcribe, $0.0045 a minute |
| Replacement for live audio | gpt-live-transcribe, $0.017 a minute |
| Earlier date to watch | gpt-4o-mini-transcribe-2025-03-20 snapshot goes on 20 January 2027 |
Is whisper being discontinued?
Yes. whisper-1 is on OpenAI's deprecations page with a shutdown date of 26 February 2027, and the 26 August changelog entry tells developers to migrate to gpt-live-transcribe or gpt-transcribe. The same notice covers all three gpt-4o transcription models, so moving from whisper-1 to gpt-4o-transcribe is not a migration; it lands you on the same deadline.
One wrinkle for teams that pinned a snapshot. A separate notice from July retires gpt-4o-mini-transcribe-2025-03-20 on 20 January 2027, five weeks before the rest. If your config names that snapshot, your real deadline is January.
What replaces OpenAI whisper-1?
gpt-transcribe replaces whisper-1 for anything already recorded. OpenAI's transcription guide splits the job in two: file transcription for a completed recording, with gpt-transcribe as the recommended model, and realtime transcription for a microphone, call or stream, with gpt-live-transcribe. Streaming output and live audio are separate decisions: you can stream the transcript of a finished file with stream=true and never open a Realtime session.
| You use today | For | Move to |
|---|---|---|
| whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe | Transcribing uploaded files | gpt-transcribe on /v1/audio/transcriptions |
| Any of the above over a live stream | Transcript deltas as people speak | gpt-live-transcribe in a Realtime transcription session |
| Committed turns over WebSocket, or detected language | Voice pipelines | gpt-transcribe in a Realtime transcription session |
| whisper-1 for timestamps, subtitles or translation | Captions, video tools, English output | No replacement named yet (see below) |
| gpt-4o-transcribe-diarize | Speaker labels | No replacement named yet (see below) |
The API shape changes a little. gpt-transcribe takes a languages list instead of the singular language field, and the request is rejected if you send both. It also accepts a free-form prompt and a keywords list for product names, drug names and account codes, where whisper-1 prompts were capped at 224 tokens. Keywords cannot contain <, > or line breaks; one bad keyword rejects the whole request, so sanitise them if they come from user data. Files are still limited to 25 MB.
The gap: timestamps, subtitles, translation and diarisation
This is the part most migration notes skip. As of 21 September 2026, the transcription guide still tells you to use whisper-1 for word timestamps, srt and vtt subtitles, and translation into English, and gpt-4o-transcribe-diarize for speaker-labelled transcripts. The file guide adds that the timestamp_granularities[] parameter is only supported for whisper-1. The gpt-transcribe model page lists the translations endpoint as not supported.
All of those models shut down on 26 February. The deprecations table lists gpt-transcribe or gpt-live-transcribe as the replacement for each, but the guides do not yet say how to get word timestamps, subtitle files, audio translation or speaker labels from them. OpenAI may document one before the date; we would not plan a release on it. If your product depends on any of those four features, isolate that code path now, watch the deprecations page, and scope a fallback you control, for example transcribing with gpt-transcribe and translating the text with a text model, before you need it.
OpenAI speech to text pricing before and after
For most whisper-1 users the move is a price cut. OpenAI's published transcription prices as of September 2026:
| Model | Status | Estimated cost |
|---|---|---|
| whisper-1 | Retiring 26 Feb 2027 | $0.006 / minute |
| gpt-4o-transcribe | Retiring 26 Feb 2027 | $0.006 / minute |
| gpt-4o-mini-transcribe | Retiring 26 Feb 2027 | $0.003 / minute |
| gpt-4o-transcribe-diarize | Retiring 26 Feb 2027 | $0.006 / minute |
| gpt-transcribe | Recommended for files | $0.0045 / minute |
| gpt-live-transcribe | Recommended for live audio | $0.017 / minute |
A worked example, assuming 100,000 minutes of recorded audio a month: whisper-1 costs 100,000 × $0.006 = $600, and gpt-transcribe costs 100,000 × $0.0045 = $450. The same volume on gpt-4o-mini-transcribe costs $300 today, so teams that chose the mini model to save money will pay 50% more on gpt-transcribe. And if you move a file workload onto gpt-live-transcribe because it sounds like the newer option, the same 100,000 minutes cost $1,700. Use the live model only for audio that is actually live.
gpt-transcribe also does not support the Batch endpoint, so there is no batch discount to fall back on for large archives.
How accurate is gpt-transcribe? Test it on your own audio
OpenAI calls gpt-transcribe high-accuracy, but the only accuracy figure that should decide your migration is the one you measure on your own recordings. OpenAI's guide makes the same point: test with representative audio and track the errors that matter to the application instead of relying only on word error rate. This is the method we use in the builds we run:
- Build a test set from production. Pull recordings that cover your real conditions: target languages and accents, code-switching, telephony and poor microphones, background noise, short utterances and long calls.
- Write reference transcripts by hand. A few hours of carefully checked references beats a large set of model output nobody verified.
- Run old and new models on the same files. Transcribe each file with your current model and with gpt-transcribe, once without context and once with
prompt,keywordsandlanguagesset. - Score word error rate after normalising case, punctuation and number formats, so you measure recognition and not formatting.
- Score the errors that cost you money separately: names, order numbers, dates, amounts, medication or product names. A model can win on word error rate and still lose here.
- Check keyword side effects. OpenAI warns that keywords are hints; confirm they improve recognition without making unspoken terms appear in the transcript.
Keep the harness. You will want it again when the next transcription model ships, and it turns "the new model feels better" into a number your team can sign off.
A whisper-1 deprecation plan that fits in one sprint
Start by searching your codebase and config for every transcription model string, including pinned snapshots, and note which features each call uses. Plain file transcription can move to gpt-transcribe now, behind a flag, once your test set passes. Live audio moves to gpt-live-transcribe; our note on what a GPT-Live-1 voice agent costs covers the neighbouring January 2027 shutdowns of the older realtime models. Anything using timestamps, subtitles, translation or diarisation goes on a watch list with an owner and a fallback.
We have run this kind of dated move before, for the GPT-4 shutdown and the Assistants API migration, and the pattern holds: the easy part moves early, and the risk sits in the one feature nobody listed. If you want a second pair of hands on the audit or the test harness, our AI engineering services team can help.
Frequently asked questions
OpenAI will remove whisper-1 from the API on 26 February 2027. The deprecation was announced on 26 August 2026 and also covers gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize. OpenAI recommends gpt-transcribe for recorded files and gpt-live-transcribe for live, streaming audio as the replacements.
gpt-transcribe replaces whisper-1 for recorded audio files on the transcriptions endpoint, at $0.0045 a minute against whisper-1's $0.006. For live microphone, call or stream audio, OpenAI recommends gpt-live-transcribe at $0.017 a minute. Timestamps, subtitles and translation have no documented replacement yet.
Yes. gpt-transcribe is listed at $0.0045 per minute and whisper-1 at $0.006, so recorded-file transcription costs 25% less after the move. Teams on gpt-4o-mini-transcribe, listed at $0.003 per minute, will pay 50% more on gpt-transcribe, and gpt-live-transcribe costs $0.017 per minute.
As of September 2026, OpenAI's guides still recommend whisper-1 for word timestamps and srt or vtt subtitles, and state that timestamp_granularities is only supported for whisper-1. Because whisper-1 shuts down on 26 February 2027, teams relying on timestamps should isolate that code path and plan a fallback.
Test gpt-transcribe accuracy on a set of your own production recordings with hand-checked reference transcripts. Run your current model and gpt-transcribe on the same files, score normalised word error rate, then separately score names, numbers and domain terms, and check that keyword hints do not insert unspoken words.
The deprecations page lists gpt-transcribe or gpt-live-transcribe as the replacement for gpt-4o-transcribe-diarize, but as of September 2026 the transcription guide still points speaker-labelled transcripts at gpt-4o-transcribe-diarize and documents no diarisation option on the new models. Treat speaker labels as an open migration item.
