According to the Google Official Blog, Google has launched the Agentic Video Understanding feature for the Gemini API. This capability enables the model to autonomously determine video analysis strategies: under traditional mode, frames are extracted at a fixed 1 FPS rate and processed in one batch; under the new mode, the model dynamically locates relevant segments along the timeline and selects visual frames, audio clips, or transcribed text based on the query, automatically increasing frame rate during fast-moving actions for re-evaluation. Google’s internal testing shows that with Gemini 3.7 Flash enabled, token consumption drops by up to 88%, analysis cost decreases by up to 66%, and accuracy improves by up to approximately 7% compared to prior versions;

the technology can pinpoint details within less than one second or extract fine-grained information from videos spanning several hours. This innovation embeds into Gemini’s core architecture what developers previously had to manually construct—a two-stage workflow of first locating key moments, then performing detailed inspection. Currently supported across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, the feature is available for both uploaded videos and YouTube videos, with no additional fees. It will be integrated into the Gemini App and YouTube’s Ask YouTube functionality in future updates.