Gemini Flash Adds Agentic Video Understanding at Up to 88% Fewer Tokens

Google DeepMind says Gemini can now analyze video by searching it rather than reading all of it. In Google's agentic video feature, the model dynamically searches, scans and inspects target segments across frames, audio and transcripts, whereas standard processing ingests video at a fixed frames-per-second rate (1 FPS by default). That lets the model spend its effort on the moments needed to answer a question.

In DeepMind's announcement chart, Gemini 3.7 Flash used 88% fewer tokens per query on two long-video benchmarks, 1H-VideoQA and LVBench (397.6K down to 47.7K, and 300.3K down to 36.0K), and 58.4% fewer on a third, Minerva, with higher accuracy in all three. The chart's footnote gives the conditions: high thinking level, low media resolution, and 1 FPS static processing as the baseline. These are DeepMind's benchmark results, not a fixed discount for every video.