OpenAI's Transcription Models Shut Down on 26 February 2027: Whisper Migration Guide

What Is Shutting Down, and When?

OpenAI is removing four transcription models from its API on 26 February 2027. On 26 August 2026 it 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.

If your business runs call analytics, meeting notes, video captions, voice-note capture or podcast transcription through OpenAI, something in that pipeline probably calls one of these models. The date is five months away, which is enough time to migrate calmly and not enough to forget about it.

A separate, earlier date applies to one pinned snapshot: OpenAI's deprecations page lists gpt-4o-mini-transcribe-2025-03-20 for shutdown on 20 January 2027, with gpt-4o-mini-transcribe-2025-12-15 as its replacement. That snapshot is then itself removed on 26 February with the rest of the family.

What Are the Replacement Models?

OpenAI released both replacements on 28 July 2026. gpt-transcribe is for completed recordings and committed turns in Realtime sessions; gpt-live-transcribe is for low-latency streaming. Both support free-form context, keyword hints and multiple expected input languages, which the older models handled less flexibly.

For most file transcription the move is cheaper, not dearer: gpt-transcribe at 0.0045 US dollars a minute undercuts whisper-1 at 0.006. Teams on gpt-4o-mini-transcribe at an estimated 0.003 will pay more per minute after migrating.

What Does the Official Guide Not Yet Answer?

This is the part of the migration most write-ups will miss. OpenAI's current file transcription guide still sends three capabilities to models on the shutdown list, and the deprecation table names only gpt-transcribe or gpt-live-transcribe as replacements.

• Word and segment timestamps. The guide says to "use whisper-1 when you need word or segment timestamps" and that the timestamp_granularities[] parameter "is only supported for whisper-1".
• Translation into English. The guide says to translate a recording into English "use /v1/audio/translations with whisper-1".
• Speaker labels. The guide says to use gpt-4o-transcribe-diarize "only when you need to identify who speaks", via the diarized_json response format.

The guide's own opening line says to use a specialised model "only if you need speaker labels, word timestamps, subtitle formats, or translation into English". As published today, it does not name which model provides those after 26 February 2027. That may change as the date approaches. Until it does, any workflow built on captions with timings, English translation or speaker separation should be treated as at risk, and the gap raised with OpenAI rather than assumed to be covered.

How Do You Migrate Plain Transcription?

For ordinary file transcription it is mostly a model-name change, with two field changes worth knowing. OpenAI's example request for gpt-transcribe:

curl https://api.openai.com/v1/audio/transcriptions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -F file="@/path/to/file/audio.mp3" \
  -F model="gpt-transcribe" \
  -F 'prompt=A customer support call about a premium plan and account AC-42.' \
  -F 'keywords[]=premium plan' \
  -F 'keywords[]=AC-42' \
  -F 'languages[]=en' \
  -F 'languages[]=fr'

• languages replaces language. OpenAI says "Don't send both fields." Older models accept one language hint as language.
• Keywords are hints, not required output. OpenAI advises including only relevant terms and checking they do not cause unspoken terms to appear.
• Character rules reject the whole request. Keywords must not contain <, >, a carriage return or a line feed, and the API rejects the request if prompt exceeds the model's length limit.
• Limits are unchanged. Files up to 25 MB in mp3, mp4, mpeg, mpga, m4a, wav and webm.

On language support, OpenAI documents the accepted code formats for gpt-transcribe (ISO 639-1, selected ISO 639-3, and regional Chinese codes) but does not publish a list of supported languages on the guide. Teams transcribing Hindi, Tamil, Marathi or mixed-language Indian calls should test accuracy on their own recordings rather than assume parity with Whisper's documented 98 languages.

What Should You Do Before 26 February 2027?

1. Inventory every call. Search code, no-code automations and vendor tools for the four model names, including tools you pay for that may use OpenAI underneath.
2. Sort by capability. Separate plain transcription from timestamps, translation and diarization. The first has a documented path; the others need a decision.
3. Test gpt-transcribe on real audio. Use your own calls and recordings, including Indian languages and code-switching if they matter to you.
4. Update request fields. Swap language for languages and add keyword hints for brand and product names.
5. Re-cost the pipeline. Most move from 0.006 to 0.0045 US dollars a minute; gpt-4o-mini-transcribe users move up from an estimated 0.003.
6. Fix pinned snapshots first. Anything on gpt-4o-mini-transcribe-2025-03-20 stops on 20 January 2027.

What Are the Common Mistakes in 2026?

• Assuming a vendor will handle it. If a captioning or call-analytics tool uses OpenAI, ask which model and what its migration plan is.
• Missing the January snapshot date. One pinned snapshot shuts down five weeks before the rest.
• Assuming timestamps and diarization migrate automatically. The published guide does not yet name replacements for them.
• Sending both language and languages. OpenAI says not to.
• Assuming Indian language accuracy. Test it; the language list for the new model is not published on the guide.
• Leaving it to February. Migrations found late tend to be found by a broken caption file.

Key Takeaways for 2026

• whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize are removed from the API on 26 February 2027.
• OpenAI recommends gpt-transcribe (0.0045 USD a minute) or gpt-live-transcribe (0.017 USD a minute), both released on 28 July 2026.
• Plain transcription migrates with a model name change and a language to languages field change.
• Timestamps, translation into English and speaker labels are still documented against retiring models, with no named replacement on the published pages.
• A pinned gpt-4o-mini-transcribe snapshot stops earlier, on 20 January 2027.
• Test Indian language accuracy on your own audio before switching production.

Distk helps growth teams across India and internationally audit which tools in their stack depend on AI models with a published shutdown date, and plan migrations before they turn into an outage. If transcription sits anywhere in your marketing or support pipeline, the inventory is the place to start.

Sources

• OpenAI, Deprecations (2026-08-26 transcription models entry).
• OpenAI developer changelog, entries for 28 July and 26 August 2026.
• OpenAI, File transcription guide.
• OpenAI, GPT Transcribe model page.
• OpenAI, GPT Live Transcribe model page.
• OpenAI API pricing.