I wanted a small Halloween post: a ghost in front of a server rack, a little fog, done. Then I thought: make the ghost dance. That was when the token monster climbed out of the rack.
One concrete example from Google's Gemini API pricing on 31 October 2026: a 1K image from Gemini 3.1 Flash Lite Image counts as 1,120 output tokens. Google counts 5,792 output tokens per second of 720p video from Gemini Omni Flash. Five seconds work out to 28,960 output tokens, roughly 26 times the counted tokens.
This is an illustration, not a benchmark of the same model: image and video use different models. Token count, compute, visual quality and money are not interchangeable units. In this example the listed output prices are about US$0.034 for the still and US$0.50 for five seconds of video at the standard rate, before extra inputs and retries.
The first ghost walks backwards. The second has seven fingers. In the third clip, the server rack turns into a fridge. Discarding one still stings; discarding a clip multiplies seconds, waiting time and budget. I now test the joke as a still first. If it does not work without motion, a camera move rarely saves it.
Use a still for a clear point, screenshot, number or infographic. Use video when motion carries the information: a sequence, a change, or a visual gag missing from one frame. This post works as an image. The token monster gets an early night. These official figures are a snapshot. Before a campaign, check the current prices and billing units, run one representative test and budget for retries. A 20-word prompt is not the total bill.
Originally published at daniele-michelin.com. More guides, free checks (SPF, TLS, headers) and an error-message lexicon: daniele-michelin.com/en/.