Google Gemini's new agent-based video analysis cuts token usage by up to 88 percent

Google is bringing an agent-based video analysis mode to several Gemini Flash models, including Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Instead of processing every frame at a fixed rate, the system decides on its own which parts of a video need closer inspection and at what resolution.

The company says this approach can reduce token usage by up to 88 percent. That is a significant efficiency gain because video is one of the most expensive input types for multimodal models, and token costs often scale with the amount of frames sampled.

The change matters because it could make video analysis more practical for real-world use cases like media indexing, surveillance, and autonomous systems. Lower token consumption means lower cost and faster responses, making it feasible to process longer clips or more video at scale.

It also reflects a broader trend of making models more agentic, letting the model decide how to allocate compute and attention rather than applying a fixed strategy. Google has not disclosed full benchmarks, but the move signals growing competition in efficient multimodal processing.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.