Gemini can now choose which parts of a video to watch, and Google says that cuts tokens by up to 88%

The numbers, and where each one comes from

• Up to 88% fewer tokens, up to 66% lower analysis cost and up to 7% better accuracy, all from Google’s announcement of 1 September 2026, measured against static processing.
• 1 frame per second: the default sampling rate the comparison is made against, adjustable through the API, from the same page.
• $0.30 in, $2.50 out per million tokens for Gemini 3.5 Flash-Lite, as our table recorded it on 5 September 2026.
• No row in that table for Gemini 3.7 Flash or 3.6 Flash on the day it was read, which is a gap in our data rather than a claim about Google’s.

Why the saving is believable, and what it is measured against

Sampling a video at a fixed rate charges you for the whole file whether your question is about the whole file or about eleven seconds of it. An hour at one frame a second is thousands of frames in the context window for every question you ask. Letting the model fetch the segments it wants replaces that with the cost of finding them, which on a long recording is a much smaller number.

The comparison, though, is against Google’s own default rather than against what a careful developer was already doing. Anyone who had already dropped the sampling rate, cut the video into chapters or pre-filtered it with a cheap pass will not see anything like the headline figure, because most of that saving was already taken.

What we will watch

Gemini 3.7 Flash and 3.6 Flash have no price row on this site. Until they appear on the pricing page we read, the practical advice is the dull kind: measure your own token counts before and after on your own footage, because the published range runs from a saving that changes your budget to one you will not notice, and where you land in it depends on the length of your videos rather than on the benchmark.