The interval decides what the model sees.
A fictional 90-second Larkspur Cloud screen recording (30 fps, 1280 by 720) walks through a data export in eight steps. Frames are picked exactly as process_video in notebook 02a picks them, and priced with OpenAI’s published image-token rule for gpt-4o. Widen the gap between frames and the short steps drop out.
- Frames sent
- Image tokens
- Steps seen, of 8
- Export finish seen
What the frames land on
One tick per frame sent
| Step on screen | From, to (s) | Length (s) | Frames on it |
|---|
The two magenta bands are the short steps: a 1.8 s confirm dialog and a 1.2 s toast. A hollow band got no frame, so the model cannot describe it. The recording and its steps are invented for this lab; the sampling and the token rule are not.
Tokens per 1280 by 720 frame
| Detail | Rule for gpt-4o | Tokens |
|---|
High detail: fit the image in 2048 by 2048, scale the shortest side down to 768 if it is longer, then count 512-pixel tiles. A 1280 by 720 frame needs 3 by 2 tiles. Source: OpenAI’s images and vision guide.