whisper-1 Is Being Retired: Six Months, and Four Transcription Models Go

What happened

The deprecation page puts it plainly: "On August 26, 2026, we notified developers using whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize of their deprecation and removal from the API on February 26, 2027." The recommended replacement for all four is gpt-live-transcribe or gpt-transcribe.

Six months is not an arbitrary number. The same page carries OpenAI's own rule: "Generally available models: At least 6 months." From 26 August to 26 February is exactly that — meaning you got the published minimum, not a day more.

Neither successor is new: both shipped on 28 July, so they had been live for a month by the time the announcement landed.

2026 JUL 28  the successors ship
2026 AUG 26  four models deprecated
             (published minimum: 6 months)
2027 FEB 26  all four shut down

What breaks

This is not a rename. Part of what whisper-1 could do is simply absent from the successor:

What you get in return: streaming on file uploads as well (whisper-1 could not do that), several language hints at once, and the detected language handed back to you.

The trap in the migration guide

This is the spine of the story, and on its own it tells you what deprecations are really like. OpenAI's own migration guide, published on 28 July, recommends in four cases that you do not migrate:

• for subtitles, "retain whisper-1 if the integration depends on native SRT or VTT output";
• for timestamps, "Timestamps: retain a model and response format that explicitly support word or segment timestamps";
• for translation, "Translation: retain /v1/audio/translations with a supported translation model";
• for speaker separation, "Speaker diarization: use gpt-4o-transcribe-diarize".

Five weeks later every one of the four models it points at was deprecated. The guide has not been rewritten since. If you opened it at the end of July and planned around it, you have to plan again now — and not because you read it wrong.

What the successor costs

Coming off whisper-1, that is a 25% cut. If you were on mini because it was the cheap one, it is a 50% rise — and there is nowhere to move, because mini has no cheaper successor.

Put a real workload on it: five hours of meeting transcripts a week, so roughly 22 hours a month. That is $7.92 a month on whisper-1 and $5.94 on gpt-transcribe; coming from mini it goes the other way, from $3.96 up to $5.94. The same hours through live, real-time transcription would be $22.44 — live is not slightly more expensive, it costs almost four times as much as processing the recording afterwards.

For scale, here is the list price of one hour of recorded audio at the larger competitors: AssemblyAI Universal-2 $0.15, Azure batch $0.18, ElevenLabs Scribe v2 $0.22, Deepgram Nova-3 $0.26, OpenAI gpt-transcribe $0.27, Google Speech-to-Text standard $0.96. So even after getting cheaper, the successor sits mid-field rather than out in front.

What about your language

Here is a gap worth naming: OpenAI publishes no language list for the new models. For whisper-1 all 98 supported languages were spelled out; for the successors you get only the format of the language code, plus a note that the API rejects a code it does not support. From the documentation alone you cannot say how well Hungarian or Spanish will hold up — only a test on your own audio can.

If that matters where you work, there is somewhere to look. Open-weight, downloadable models that name Hungarian explicitly: NVIDIA parakeet-tdt-0.6b-v3 (600M) and canary-1b-v2 (1B), both under CC BY 4.0, covering 25 European languages including Hungarian — and Canary translates as well. Mistral Voxtral lists eight languages, Hungarian not among them. Among hosted providers, ElevenLabs Scribe places both Hungarian and Spanish in its best band, under 5% word error rate; AssemblyAI puts Spanish in the best band and Hungarian in the second.

And the part that settles most nerves: downloadable Whisper is not going anywhere. "Whisper's code and model weights are released under the MIT License." The retirement touches only OpenAI's hosted endpoint. In fact the hosted whisper-1 was built on Whisper V2 by OpenAI's own account, while the publicly downloadable line already includes large-v3 and the faster turbo. Anyone who was paying purely for the convenience gets a newer generation by running it themselves.

What to do now

1. Find every place those four model names appear. Not only in code: in automations, in no-code flows, and in the settings of note-taking and subtitling tools. (More: handling model retirements in your automations.)
2. Write down which formats you actually use. If SRT or VTT feeds a video editor downstream, or if search inside a transcript rests on word-level timestamps, this is not a price change for you but lost functionality — and you have six months to replace it.
3. If you separate speakers, plan for it now. There is no named successor. It is either another provider, an open-weight model, or a separate diarisation step.
4. Measure, don't assume. Take a ten-minute recording that is typical of your own material — in your language, with the usual noise and jargon — and run it through the old model and the successor. Word error rate has to be read off your own audio, because no published figure exists.

More on the subject: transcription and subtitles with AI, and on what to do in general when a model of yours is retired: when your model is deprecated — the general steps are there too.

Common misconceptions

• "Whisper is being shut down." No. The hosted whisper-1 endpoint is. Open-weight Whisper stays downloadable under the MIT licence, and in newer versions than the hosted one.
• "The successor does everything the old one did." It does no subtitle formats, no word-level timestamps, no translation and no speaker separation — at least not according to the documentation.
• "It gets cheaper for everyone." For anyone coming off mini the list price rises by half.
• "February is ages away." Migration here is not a model-name swap: if a format or a timestamp goes with it, that reaches into the processing chain. Nobody rewrites that in the final week.
• "Live transcription costs about the same." Almost four times as much. If you do not need the text instantly, processing the recording afterwards is by far the cheaper route.

What to take away

A retirement is not news in itself: every provider does it, regularly. The lesson in this case is rather that notice time is not the same thing as preparation time. You have six months to swap the model — but if a format, a timestamp or an endpoint goes with it, you are not swapping, you are redesigning. Which is why it pays to go through what your process actually rests on now, and not once the old model starts returning errors. The practical handhold is the same as always: know which model identifier runs where in your setup, and keep a tested plan B for every step that matters.