The unit of work is a 30-second window.

Whisper's encoder takes exactly 30 s of 16 kHz mono audio as a log-mel spectrogram — 480,000 samples, 3,000 mel frames at a 10 ms hop — and it pads anything shorter with zeros. So a 4-second voice memo costs a whole window: 30 s of encode, 7.5× its own length. And one extra second past 30 buys 30 more.

The engine is free and needs no account. Silence costs exactly what speech costs, because the mel of a silent window is the same shape as any other — there is no discount to find, only audio to remove. Which is the trap: removing it is the one thing that lowers the bill, and it puts the model on the trimmed clock, so every timestamp comes back early by all the silence removed before it — up to 18 minutes at the end of an hour that was 30% silence. And the answer has a resolution you did not choose: 1,500 encoder positions over 30 s makes a position 20 ms, so every timestamp the model can express is a multiple of 0.02 s. At 25 fps that divides the frame and every frame boundary is expressible; at 24 fps one in twelve is; at 30000/1001 it is one in six hundred. Finally the hosted ceiling is a duration: 25 MiB at the rate the model actually consumes is 13:39 of audio, whatever the format. Every lane ships a worked example that costs nothing. Writing the report needs credits.