Google's Gemini Now Picks the Key Moments in Videos for You

Google has made Gemini, its AI model, smarter about watching videos. Instead of processing every single frame at a fixed speed, Gemini can now act like a student flipping through a textbook: it searches transcripts, audio, and images to find the part that matters, then zooms in for detail. Google calls this "agentic video understanding," and it says the approach can cut the amount of data processed by up to 88% and reduce analysis costs by up to 66%, while slightly improving accuracy.

Why should you care? Cheaper and more efficient video analysis means AI tools can handle much longer recordings without racking up huge bills. Google points to examples like 90-minute lectures, 10-minute how-to videos, and multi-hour monitoring footage. Instead of paying to process every frame, developers only pay for the parts Gemini decides are relevant. That could lead to cheaper apps for searching videos, editing clips, detecting anomalies, or counting repeated actions.

The feature is available now through the Gemini API, and it supports several versions of Gemini as well as uploaded videos and public YouTube videos. More interesting for everyday users: Google says this technology will power the "Ask YouTube" feature on video watch pages in the coming months. That means you could ask a question about a video and get answers grounded in what actually appears in it, without watching the whole thing.

There is one honest limitation. Google recommends the old, frame-by-frame approach for shorter videos, especially clips under five minutes, because full coverage keeps latency low and ensures nothing is missed. So while Gemini's new skill is a big step toward AI that decides how to look at the world, it's not always the best tool for every job.