Gemini Gets Smarter at Video: 88% Fewer Tokens

Google DeepMind has launched agentic video understanding for Gemini: instead of scanning the entire file, the model now selects the frames, playback speed, and modality — visual, audio, or subtitles — based on the query. Token consumption dropped 88%, analysis cost fell 66%, and accuracy improved 7%.

It's now possible to find a specific moment in an hours-long video in a fraction of a second, automatically count repeated actions, or detect an anomaly. The feature is available in Google AI Studio and through the Gemini API at standard rates, with no additional markup.