Transcript endpoints return the platform's captions when they exist. source is captions or auto_captions, and the price is 1 credit. If there are no captions, the audio can be transcribed. Then source is ai, with a language and timed segments. Check source when the difference matters.
AI transcription costs 1 credit for up to 2 minutes of audio, then 1 more credit for each started minute. You pay only when the transcript is delivered. A failed transcription is free. A repeat you already paid for is free with max_age_days.
Set ai_fallback=false to skip audio. The endpoint then returns nothing and charges nothing. The transcription guide lists which transcript endpoints accept it. use_ai_as_fallback is accepted as the same parameter.
Short clips finish in the normal request time. A call still running after about 25 seconds answers 202 with a job to poll. Polling is free. See batch and jobs.
If the lane is busy, or the team's daily allowance is used up, the error is overloaded (retryable, free). Media longer than about 4 hours, or with no speech, answers transcript_unavailable.
Transcript from any video URL picks the endpoint for you and uses the same price and the same cache.
Full detail: AI transcription.