Google just gave the Gemini API a new trick: Agentic Video Understanding. Instead of blindly sampling one frame per second and shoving it all into context, the model now decides which parts of a video actually matter — then watches them the right way. It can scan along the timeline, pick out relevant video, audio, or transcript depending on the question, and even bump up the frame rate for fast motion to take a closer look.
Google says its own tests show Gemini 3.7 Flash with this feature uses up to 88% fewer tokens, cuts analysis costs by up to 66%, and delivers a relative accuracy boost of roughly 7%. It can pinpoint a moment shorter than one second — or find a single detail buried in hours of footage.
Technically, Google took the whole "find it first, then zoom in" workflow developers used to build themselves and baked it straight into Gemini. It's available now in Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, for both uploaded videos and YouTube videos, with no extra feature fee. Down the line, it's also coming to the Gemini app and YouTube's Ask YouTube.
Source: Google blog