OpenAI's recommended transcription model is a quarter cheaper — and cannot make subtitles

OpenAI released two transcription models on 28 July 2026: GPT Transcribe for completed recordings and GPT Live Transcribe for live audio. Since then the company's own transcription guide points developers to them first, and describes the earlier models — gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-realtime-whisper — as fine for existing integrations but "not the recommended starting models". The headline is the price. GPT Transcribe bills $0.0045 per minute of audio, a quarter below the $0.006 charged for whisper-1 and gpt-4o-transcribe. It also takes richer instructions than its predecessors: a free-form description of the recording, keyword hints for names and jargon, and a list of expected input languages instead of a single one — the older models accept the singular `language` field and assume one tongue per file. Read the same guide further down, though, and the recommendation comes with a list of exceptions. Word-level timestamps and srt/vtt subtitles: whisper-1. Speaker-labelled transcripts: gpt-4o-transcribe-diarize. Translating a recording into English: whisper-1 again. All three of those models bill $0.006 a minute. Anyone producing subtitles, court-style transcripts or translated recordings therefore pays 33% more than the recommended rate and does so on a model the vendor has stopped recommending — a combination that usually precedes a retirement notice, though none has been announced for any of them. The second number worth noting sits between the two new models. Live transcription costs $0.017 a minute against $0.0045 for the same audio submitted as a file: 3.8 times as much, with immediacy the only difference. That live rate is unchanged from gpt-realtime-whisper, the streaming model it supersedes. In other words the July release cut the price of unhurried transcription by a quarter and left the price of impatience exactly where it was. Across OpenAI's current speech-to-text line-up the same minute of audio can cost anywhere from $0.003 (gpt-4o-mini-transcribe) to $0.017 (either streaming model) — a spread of 5.7 times, and the cheapest option is not among the recommended ones either.