API
OpenAI
OpenAI has deprecated its older speech-to-text API models — whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize — and will remove them from the API on 26 February 2027
departs 26 Feb 2027
Quick answers
- Is it true that OpenAI deprecates whisper-1 and its gpt-4o-transcribe speech-to-text models?
- Yes. OpenAI has officially announced it, effective 26 Feb 2027. On 26 August 2026 OpenAI notified developers that four transcription (speech-to-text) models are deprecated and will stop responding in the API on 26 February 2027: whisper-1 (the long-standing Whisper transcription endpoint), gpt-4o-transcribe and gpt-4o-mini-transcribe (the GPT-4o-based transcription models), and gpt-4o-transcribe-diarize (the variant that labels who is speaking).
- When does it take effect?
- 26 Feb 2027. Verified against the vendor's own announcement.
- What should I use instead?
- OpenAI points every retired transcription model to the same two replacements: gpt-live-transcribe or gpt-transcribe (see OpenAI's transcription guide).
What this means for you
On 26 August 2026 OpenAI notified developers that four transcription (speech-to-text) models are deprecated and will stop responding in the API on 26 February 2027: whisper-1 (the long-standing Whisper transcription endpoint), gpt-4o-transcribe and gpt-4o-mini-transcribe (the GPT-4o-based transcription models), and gpt-4o-transcribe-diarize (the variant that labels who is speaking). Deprecated means they still work today; on the shutdown date, any API call that names one of these models starts failing. Anything that hardcodes one of these model IDs — a meeting-transcription pipeline, a voice-note or podcast transcriber, a call-center analytics job, an accessibility captioning feature — will error on 26 February 2027 unless the model string is updated first. This is the transcription counterpart to OpenAI's legacy audio and realtime deprecation (shutdown 20 January 2027); note that whisper-1 here is the hosted OpenAI API model, not the open-weight Whisper you can run yourself, which is unaffected.
Where to go
OpenAI points every retired transcription model to the same two replacements: gpt-live-transcribe or gpt-transcribe (see OpenAI's transcription guide). Find every place your code or config pins one of the old model IDs — search your codebase, prompt store, and any environment/config for the exact strings whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize — point them at gpt-live-transcribe or gpt-transcribe, and re-run your transcription tests. Newer models can shift accuracy, latency, timestamp/segment formatting, and token accounting even when the request shape is unchanged, and diarization behaviour in particular is worth re-validating if you relied on gpt-4o-transcribe-diarize. Teams that would rather not track OpenAI's model cadence can self-host open-weight Whisper or evaluate another speech-to-text provider, but that is a larger migration than swapping a model string.
- gpt-live-transcribe — OpenAI's named transcription replacement
- gpt-transcribe — OpenAI's named transcription replacement
- Self-hosted open-weight Whisper, or another speech-to-text provider — a larger migration off OpenAI's hosted models
OpenAI is retiring its older speech-to-text models. On 26 August 2026 the company notified developers that four transcription models are deprecated and will be removed from the API on 26 February 2027: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize.
“Deprecated” is the warning, not the cutoff. The models keep working through the runway; on 26 February 2027 an API call that names any of them starts failing. If you transcribe audio through OpenAI’s API today, the question is simply whether one of these model IDs is written down anywhere in your stack.
Who this hits
Anything that pins one of these models by name breaks on the date: a meeting- or interview-transcription pipeline, a podcast or voice-note transcriber, call-center or sales-call analytics, an accessibility captioning feature. If your code selects a transcription model from a variable and you can retarget it in one place, this is a small change; if the model ID is scattered across services, scripts, and config, the work is finding every copy.
One clarification worth making: whisper-1 here is the hosted OpenAI API transcription model. The open-weight Whisper you can download and run on your own hardware is a different thing and is not affected by this deprecation.
What to move to
OpenAI points all four retired models at the same two replacements: gpt-live-transcribe or gpt-transcribe, per its transcription guide. The migration is mostly a rename — swap the model string — but re-run your transcription tests afterward. A newer model can change accuracy, latency, and the shape of timestamps and segments even when the request looks identical, and if you relied on gpt-4o-transcribe-diarize for speaker labels, re-validate that diarization still behaves the way your product expects. Teams that would rather not follow OpenAI’s model cadence can self-host open-weight Whisper or evaluate another speech-to-text provider — a larger move than editing a model ID, but it takes the cadence question off the table.
This entry tracks the official OpenAI API deprecations page; if OpenAI moves the date or changes the recommended replacements, the entry will be updated with a changelog note.
Entry changelog
- — last reviewed.
- — Entry created from OpenAI's official API deprecations page. Announced 26 August 2026; shutdown 26 February 2027. Affected models (whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize) and the recommended replacements (gpt-live-transcribe, gpt-transcribe) are taken directly from the deprecations table. Four text: fingerprints match the retiring model IDs only; the replacements gpt-live-transcribe / gpt-transcribe and the surviving gpt-4o-mini-transcribe-2025-12-15 snapshot are boundary-safe and deliberately not fingerprinted.
Put this countdown in your README
Building on OpenAI? Drop this badge into a project’s README or docs — it shows the days left and
refreshes on its own every night from this page. 
[](https://lastcall.dev/entries/openai-transcription-models-eol)Spot an error or a changed date? Tell us and it’s corrected the same night.