≈

The passage around this formula

where, at default resolution, the sampling rate f is one frame per second, each frame costs $\tau_{\text{frame}}$ $\approx$ 258 tokens, and the audio track costs $\tau_{\text{audio}}$ $\approx$ 32 tokens per second — giving $c_{\text{video}}$ $\approx$ 290 tokens per second, close to the approximately 300 tokens per second Google documents directly for default-resolution video [ 5 ] . The formula is unremarkable arithmetic, but what it exposes is the real assumption underneath the “native” claim: a single request is billed, and therefore presumably processed, as one combined audio-visual stream, not as a picture track with a separately bolted-on transcript.

Read this part in the article →