XNews

Google Gemini Flash Agentic Video Understanding: 88% Token Cut, 7% Accuracy Gain

Google Gemini Flash Agentic Video Understanding: 88% Token Cut, 7% Accuracy Gain

Google introduced Agentic Video Understanding for Gemini Flash models, shifting from passive fixed-frame ingestion to an agentic loop where the model decides where to look, at what speed, and which modality (frames, audio, subtitles). This reduces token consumption by up to 88%, analysis cost by up to 66%, and improves accuracy by up to 7%. Four key capabilities: sub-second moment localization, long video needle-in-haystack search, anomaly detection with dynamic sampling, and action/object counting via variable-speed replay. The approach extends agentic vision techniques from images to the temporal dimension of video.

Source posts