Frame budget

The cost structure is simple. A transcript for a ten-minute video costs a few thousand tokens. The frames are everything — 80 of them at 512px wide runs 50–80k tokens, and raising the resolution to 1024 roughly quadruples the per-frame cost.

So there is a ceiling: 100 frames total, 2 fps. Within that, the budget is allocated by duration.

The problem is visible there. The budget grows about threefold, from 30 frames to 100. The length grows sixtyfold. Hand it a thirty-minute video whole and you are looking at one frame every eighteen seconds. Everything in between is known only from the transcript.

Even with a transcript, you need the frames

“Wouldn’t the transcript be enough?” was my own first thought too. Transcripts are cheap and frames are expensive.

It isn’t enough. Too much never gets said out loud. Screen layout, which button sits where, which way a graph bends, what actually moves in the demo — the presenter says “as you can see here” and moves on. What survives in the transcript is the word “here.”

And for someone watching in order to build, the “here” is usually the part that matters. So the frames get bought, expensive as they are.

Which is why long videos get cut

Sampling a long video sparsely is almost always worse than sampling the part you need densely. Name a range and the whole budget goes inside it — 2:15 to 2:45 at 3 fps puts 90 frames into those thirty seconds.

Same 90 frames, sixtyfold difference in coverage. Spread across thirty minutes it’s one frame every twenty seconds; concentrated into thirty seconds it’s three per second.

So a long video gets a question first: which part. The most wasteful tokens are the ones burnt before deciding that. And past ten minutes a warning prints, which the answer is required to repeat — because having sampled sparsely is itself part of the result.

The same video is never watched twice

A follow-up question in the same session does not re-run anything. The frames and transcript are already in context; extracting them again is paying the same cost twice.

There’s a reason that has to be written down as a rule. From the calling side, “new question, new run” feels natural — so leave it unsaid and the second question spends another 80k tokens.

The resolution stays where it is

The default is 512px wide. At 1024 the per-frame token cost roughly quadruples, which puts 80 frames at 200–300k tokens.

There is exactly one case worth raising it for: needing to read small text on screen. And even then, only that range goes up. Full-length high resolution makes the entire budget design meaningless.