Google adds video Agent to Gemini, cutting costs by up to 66%

Beating AI News: Google has launched Agentic Video Understanding for its Gemini API. The feature enables the model to autonomously decide which video segments to process and how to analyze them. The standard mode defaults to extracting frames at a fixed 1 FPS rate before feeding all frames into the model’s context window at once. The new mode, by contrast, automatically identifies relevant segments along the video timeline, selecting visual frames, audio, or transcribed text based on the query. When fast-paced actions are detected, it can even increase the frame rate to re-examine those segments. Google’s internal testing shows that enabling this feature on Gemini 3.7 Flash cuts token consumption by up to 88%, analysis costs by up to 66%, and boosts relative accuracy by up to around 7%. The capability can pinpoint moments shorter than one second and locate specific details within hours-long videos. Technically, this integrates the "localize first, refine later" workflow that developers previously had to build themselves directly into Gemini’s core. The feature is now supported on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. It works with both uploaded videos and YouTube videos, with no extra fees for the functionality. Future plans include integrating it into the Gemini App and YouTube’s Ask YouTube tool.