OpenAI’s pay-as-you-go transcription looks straightforward: longer recordings mean more tokens and higher bills. Developer George Mandis stumbled on a way to shrink both time and cost dramatically by literally making the minutes shorter.
A simple speed hack
Mandis fed his audio through FFmpeg’s atempo filter, tripling playback speed without pitch distortion. The accelerated track remained clear enough for Whisper to understand every sentence while its runtime collapsed from 25 to barely 8 minutes.
Why faster costs less
Because billing is pegged to audio duration, a 3x-sped file burns roughly one-third of the tokens and one-third of the wall-clock time. In his test, charges fell from 0.24 USD to 0.07 USD at 2x and just 0.04 USD at 3x—an instant 83 % saving.
How to try it
Download the talk with yt-dlp, then run ffmpeg -i talk.mp3 -filter:a "atempo=3.0" talk_3x.mp3. Upload the new file to ChatGPT and watch the transcript arrive in a fraction of the usual time. Going beyond 3x, however, degrades clarity and risks transcription errors.