Gemini video analysis cuts token use by 88%

This update refines Google's earlier move into "agentic vision," which shipped for Gemini 3 Flash in January. That feature let the model write and run Python to edit images, checking results in a loop. Now the same principle applies to video. The model decides which parts of a video to inspect, through frames, audio, or transcript, and retrieves only what it needs, cutting token use by up to 88% on benchmarks like 1H-VideoQA and LVBench.

The practical payoff is accuracy on small details, such as state changes or cuts shorter than one second, which static per-second sampling could miss. Google also reports that on LongVideoBench, Gemini 3.7 Flash leads in overall quality and cost efficiency, suggesting the agent-based approach is not just cheaper but often more precise.

Google says the feature will reach all Gemini app users on Flash and Flash Lite soon, with "Ask YouTube" integration following over the coming months. Because the API charges standard token rates with no extra fee, the cost savings could make sophisticated video analysis more accessible to developers.