OpenAI retires whisper-1 and the gpt-4o-transcribe line

OpenAI announced on August 26 that four transcription models — whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize — are deprecated and shut off from the API on February 26, 2027. Six months is a generous window by 2026 standards. It is also six months that will vanish, because speech-to-text is the layer nobody remembers they built: it sits inside the voicemail summarizer, the call-notes job, the podcast pipeline, the intake form that takes a voice memo.

What actually happened

Per OpenAI's deprecations page, the announcement date is 2026-08-26 and the shutdown date is 2027-02-26 for all four models. The recommended replacements are gpt-live-transcribe and gpt-transcribe.

whisper-1 is the notable one. It was the original hosted Whisper endpoint, it has been in production code since 2023, and it is the string most likely to be sitting hardcoded in a script somebody wrote once and never touched again. gpt-4o-transcribe-diarize matters for a different reason: if you relied on it for speaker separation, verify the replacement handles diarization the way your downstream code expects before you swap the model name and call it done.

Why model deprecation matters for your business

The migration itself is usually a one-line change. The risk is everywhere else.

Find every call site first. Grep your repos for all four model strings, then check the places code doesn't live: n8n and Zapier steps, cron jobs on a box nobody logs into, a Google Apps Script, a vendor integration you configured in a web UI two years ago. The API call that breaks on February 27 will be the one that isn't in git.

Re-benchmark, don't just re-point. A new transcription model has a different word error rate on your audio — your accents, your industry jargon, your bad phone lines, your product names. Pull 20 to 30 real recordings that look like your actual traffic, transcribe them on both the old and new model, and diff. If you have never measured WER on your own audio, this deprecation is the excuse to start, because you cannot defend a number you have not measured.

Check what reads the output. If anything downstream parses timestamps, segments, confidence scores, or speaker labels, the response shape is where the breakage hides — not the transcription quality.

Price it before you're forced to. You are re-selecting a vendor anyway. That is the moment to check whether OpenAI is still the right call for this specific job, or whether a local Whisper deployment or a competing API fits your volume better. Deprecations are cheap opportunities to stop paying rent on a default you picked in 2023.

Do the audit in September while the clock is generous. February migrations done in February are how a working phone-intake system stops working on a Monday.

Sources: OpenAI API deprecations, OpenAI API changelog.