OpenAI whisper-1 shutdown: migration guide and alternatives

OpenAI plans to remove whisper-1 from its API on February 26, 2027. It told developers on August 26, 2026, and the same notice covers gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize. For all four, OpenAI names gpt-transcribe (recorded files) or gpt-live-transcribe (live audio) as the replacement (OpenAI deprecations). If you only need plain transcription, switching is mostly a one-line model change on the same endpoint. The hard part is everything else whisper-1 does: SRT and VTT subtitles, word and segment timestamps, and translation into English. OpenAI's migration guide says to keep whisper-1 for those, and as of September 30, 2026 its docs name no replacement for them. Only the hosted API model is affected. The open-source Whisper model stays free to download and run.

What's being removed, and when

Prices are OpenAI's per-minute estimates from its pricing page, checked September 30, 2026. The two dates to know:

• August 26, 2026: OpenAI "notified developers using whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize of their deprecation."
• February 26, 2027: the planned removal from the API. Assume any request that still names one of these models will stop working that day.

The replacements: gpt-transcribe and gpt-live-transcribe

OpenAI released both models on July 29, 2026 (OpenAI Developer Community).

• gpt-transcribe is for completed audio: uploaded files, streamed transcripts of a finished file, and committed turns in a Realtime session over WebSocket. It costs $0.0045 per minute (model page). OpenAI's speech-to-text guide calls it "the recommended model for transcribing recorded speech in its original language."
• gpt-live-transcribe is for audio that is still arriving, like a microphone, a call or a live stream. It runs in Realtime transcription sessions over WebSocket or WebRTC and costs $0.017 per minute (model page).

The file limits don't change. Uploads are capped at 25 MB, and the accepted formats are mp3, mp4, mpeg, mpga, m4a, wav and webm (OpenAI speech-to-text guide).

How to migrate a plain transcription call

OpenAI's migration cookbook keeps the endpoint (POST /v1/audio/transcriptions) and the file upload. What changes is the model name, some request fields and the response.

1. Change the model. Replace whisper-1 with gpt-transcribe.
2. Read the new response. You get JSON with text and a languages array of detected languages. An empty array ("languages": []) is a valid result that means the model couldn't make a reliable prediction. It isn't an error.
3. Swap language for languages. The new model takes an array of expected languages. Don't send both fields.
4. Move vocabulary into keywords. Use prompt for free-form context about the recording and keywords for literal terms you expect to hear, such as product names or account IDs. Each keyword must be a single line without <, > or line breaks. One invalid keyword makes the API reject the whole request.
5. Turn on streaming if you want partial text. Set stream=true with gpt-transcribe to receive transcript.text.delta events while it works. whisper-1 ignores stream.
6. Test on your own audio. The cookbook suggests comparing word error rate plus exact-match accuracy on names, numbers and email addresses. It also suggests checking that hinted keywords don't appear in the transcript when nobody said them.

A minimal request looks like this:

1curl https://api.openai.com/v1/audio/transcriptions \
2  -H "Authorization: Bearer $OPENAI_API_KEY" \
3  -F file="@meeting.mp3" \
4  -F model="gpt-transcribe" \
5  -F 'languages[]=en' \
6  -F 'keywords[]=AC-42' \
7  -F 'prompt=A customer support call about a premium plan.'

What has no drop-in replacement

The cookbook's compatibility section warns that some whisper-1 features "do not have a drop-in replacement in gpt-transcribe or gpt-live-transcribe." Here's where each one stands in OpenAI's docs as of September 30, 2026:

Building your own subtitles from gpt-transcribe isn't simple either. Its documented response contains text and detected languages, but no timestamps, so there's nothing to time the captions against. OpenAI's pricing page also lists a live translation model, gpt-realtime-translate ($0.034 per minute). It's a Realtime model, not a stand-in for the file-based translations endpoint.

OpenAI may add these features before the deadline. Check the deprecations page and the speech-to-text guide before you rebuild anything.

Options for subtitles, timestamps and translation

1. Keep whisper-1 until the deadline

It still works, and the cookbook recommends it for these jobs. It buys time, not a solution: plan a fallback before February 2027.

2. Run open-source Whisper yourself

The deprecation covers API models. Whisper itself stays on GitHub under the MIT License. OpenAI's API reference says whisper-1 "is powered by our open source Whisper V2 model," so self-hosting large-v2 keeps you on the same model family. The command-line tool writes txt, vtt, srt, tsv and json, and it has experimental word timestamps. It translates into English with the multilingual models; the default turbo model isn't trained for translation, so use medium or large. The costs are your own hardware, a GPU for any real speed, and the upkeep. Our OpenAI Whisper guide covers setup and model choice.

3. Switch to another speech-to-text API

Two examples with subtitle output, per their own docs and pricing pages (checked September 30, 2026):

• AssemblyAI has endpoints that return SRT or VTT for a finished transcript, with an optional characters-per-caption limit. Pre-recorded pricing starts at $0.15 per hour for Universal-2 (about $0.0025 per minute). Universal-3.5 Pro costs $0.21 per hour, and speaker labels add $0.02 per hour (AssemblyAI pricing).
• Deepgram returns JSON with timestamps. Its open-source caption packages for JavaScript and Python convert that JSON into WebVTT or SRT. Nova-3 costs $0.0043 per minute for one language and $0.0052 per minute multilingual, with speaker diarization included on pay-as-you-go (Deepgram pricing).

We haven't compared their translation features or accuracy, so test them on your own audio. For a wider field, see our transcription software comparison.

4. If you called whisper-1 only to caption your own videos

If the API was just a quick way to get transcripts or subtitles for your own recordings, a hosted editor may be less work than a migration. Disclosure: we make Exemplary AI. It transcribes audio and video in 99 languages, lets you fix the text in an editor, exports SRT, VTT or TXT files, and can translate the result. If you're not sure which subtitle format you need, see SRT vs VTT.

What switching costs

Here's what 1,000 hours of audio (60,000 minutes) costs at OpenAI's list prices:

Moving from whisper-1 to gpt-transcribe cuts the per-minute price by a quarter. Moving from gpt-4o-mini-transcribe raises it by half. Moving batch jobs to gpt-live-transcribe makes little sense: it's built for live audio and costs almost four times as much.

A migration checklist

1. Find every call. Search your code and config for whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize.
2. Label each one by what it needs: plain text, subtitles, timestamps, translation, speaker labels or live audio.
3. Move the plain-text calls to gpt-transcribe now. Change the language and keyword fields and test them with your own audio.
4. Move live captioning to gpt-live-transcribe. Keep your existing Realtime session code, and tune the model's delay setting to trade speed against accuracy.
5. Choose a path for the rest: keep whisper-1 for now, self-host Whisper, or change providers.
6. Pick an internal deadline weeks before February 26, 2027, and recheck OpenAI's deprecations page as it gets closer.

FAQ

When is OpenAI shutting down whisper-1?

OpenAI plans to remove whisper-1 from the API on February 26, 2027. It announced the deprecation on August 26, 2026.

Is open-source Whisper being discontinued?

No. The notice lists hosted API models only. The open-source Whisper code and weights remain on GitHub under the MIT License.

What replaces whisper-1?

OpenAI names gpt-transcribe for recorded audio and gpt-live-transcribe for live audio. Both use the existing transcription endpoints.

Does gpt-transcribe support SRT, VTT or timestamps?

Not according to OpenAI's docs as of September 30, 2026. The migration guide says to keep whisper-1 for native SRT and VTT output, and timestamp granularities are documented only for whisper-1.

Is gpt-transcribe cheaper than whisper-1?

Yes. It costs $0.0045 per minute, compared with $0.006 for whisper-1. It is more expensive than gpt-4o-mini-transcribe at $0.003, which is also being removed.