On September 1, Google launched Active Video Understanding, allowing models such as Gemini 3.7 Flash to decide on their own which segment to watch, how fast to watch it, and to call internal tools to closely read clips. Official benchmarks show that when enabled, token consumption drops by up to 88%, analysis costs fall by up to 66%, and accuracy improves by up to 7%. As multimodal AI moves from frame-by-frame image viewing to on-demand understanding, its cost structure changes accordingly.

Over the past six weeks, Google has released three Flash models, but what truly deserves attention is Active Video Understanding, launched on September 1.

t lets AI decide for itself which segment of the picture to focus on, when to rewind and watch again, and how to watch—rather than reading frame by frame at a fixed frequency.

In an official demo, the model was asked to count applause—something traditional frame-by-frame methods struggle with, because