OpenAI's recommended transcription model is not its cheapest — the one it stopped recommending costs a third less
Published: 8/16/2026 · Source: OpenAI — API pricing ↗
Since 28 July 2026 OpenAI has told developers to start with GPT Transcribe when they need speech turned into text. Its pricing page, read today, shows that this recommendation and the cheapest option are not the same thing.
GPT Transcribe is billed at $0.0045 per minute of audio. GPT-4o mini Transcribe — released back in March 2025, still on sale, no longer recommended — does the same job for $0.003 per minute. That is a third less for every hour of recording, and the gap has been open since the recommendation changed in July.
The full list of what OpenAI charges per minute of audio, taken from its own pricing page:
| Model | Price per minute |
| --- | --- |
| GPT-4o mini Transcribe | $0.003 |
| GPT Transcribe (recommended) | $0.0045 |
| GPT-4o Transcribe | $0.006 |
| GPT-4o Transcribe Diarize | $0.006 |
| Whisper | $0.006 |
| GPT Live Transcribe | $0.017 |
| GPT-Realtime-Whisper | $0.017 |
| GPT-Realtime-Translate | $0.034 |
Between the cheapest and the most expensive entry the spread is 11.3x — though the top of the table buys something different: live transcription as the words are spoken, and in the last row live translation, rather than a file processed after the fact.
Within the batch tier, where the comparison is like for like, the recommendation costs 50% more than the floor. What the extra buys is real and OpenAI documents it: GPT Transcribe accepts a free-form description of the recording, a list of keyword hints, and a list of expected input languages rather than a single one, which matters for recordings that switch language mid-sentence. It also streams output before the file has been processed in full. Whether that is worth 50% depends entirely on the recording — for clean single-language audio, it is money spent on features that will not fire.
One more asymmetry surfaced while we were checking the model cards. GPT-4o mini Transcribe carries two dated snapshots, so a customer can pin the exact version being billed; its twice-as-expensive sibling GPT-4o Transcribe has a single undated alias and no pinnable version at all. The cheaper model is the better-governed one.
On 20 July 2026 OpenAI placed the original March 2025 snapshot of the mini model on its deprecation list, with a shutdown date of 20 January 2027 and the December 2025 snapshot named as the replacement. The models themselves carry no shutdown date.
Both GPT-4o transcription models now have profiles in the wujec.ai catalogue.