Gemini learns to decide for itself which video segments to watch, with 88% fewer tokens
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
SI watches video the way a person does, skipping ahead to what matters, which makes analyzing multi-hour recordings fast and cheap.
In "agentic video" mode, the Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite models do not scan a clip frame by frame at a fixed rate but find and study the relevant segments themselves. This cuts token use by up to 88% and cost by up to 66% and raises accuracy by up to 7%; the mode is suited to finding moments with sub-second precision in multi-hour recordings and to counting objects and actions. The feature is already available in the API and will later come to the Gemini app and "Ask YouTube".
Sources 1