Whisper shuts down on 26 February. The risk is not the deadline.

On 26 August 2026 OpenAI deprecated whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, with removal from the API on 26 February 2027. The named replacements are gpt-transcribe and gpt-live-transcribe.

It arrives alongside a related date. The legacy audio and realtime family, including gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime and the mini variants, shuts down on 20 January 2027, announced 20 July 2026. If you do anything with speech, you likely have exposure to both.

Whisper has one of the largest install bases of any OpenAI endpoint, and much of that install base is glue code. Transcription is typically not a feature in itself: it is the first step of a pipeline that summarises a call, indexes a meeting, generates subtitles or feeds a search index.

That shape is what makes it risky. Pipeline steps run on a schedule or off a queue, so when they start failing the first symptom is usually an absence rather than an error: summaries stop appearing, a search index quietly stops growing, a metric drifts. Nobody is watching a screen at the moment it breaks.

Five months of notice is genuinely generous. The failure mode is not that the window is too short, it is that a comfortable window invites deferral until the window is no longer comfortable.

The replacement path is named, which puts this in the easier category of deprecation: a model string change plus verification. The verification is the part worth budgeting for, because transcription quality is not uniform across accents, audio quality, domain vocabulary and language, and your evaluation set probably does not cover all four.

Note that gpt-live-transcribe and gpt-transcribe are different tools for different shapes of work. If your pipeline processes a completed file, you want the batch-style endpoint; if you are transcribing an ongoing stream, you want the live one. Picking by name similarity rather than by workload is the easy mistake here.