Google introduced agentic video understanding on September 1, 2026, across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. The company says the feature improves accuracy while cutting token consumption by up to 88 percent and cost by up to 66 percent, with quality gains of up to 7 percent. It is available now for uploaded files and YouTube links through the Gemini API in Google AI Studio and on the Gemini Enterprise Agent Platform.

The pitch is that static video analysis wastes tokens. Today’s default path feeds a model frames at a fixed rate, often one frame per second. Agentic video understanding lets Gemini use its own video tools the way agentic vision already mixes code execution with image understanding. Google says that loop unlocks sub-second moment retrieval, tighter anomaly detection, and more precise counting, including cases where the model rewatches a clip at different frame rates.

Developers turn the mode on by setting processing to “agentic” in the API. Google is charging standard Gemini token prices with no extra feature fee.

From the API into consumer products

The same efficiency story is headed at consumer surfaces. Google says agentic video understanding will roll out to all users in the Gemini app on Flash and Flash-Lite models soon. In the coming months it will also sit behind YouTube’s Ask YouTube feature on the watch page, where Gemini is supposed to answer questions grounded in the visuals rather than the title and description.

Google’s own examples include counting fast motion by scanning and rewatching at different frame rates, and answering complex questions about a long video while using far fewer tokens than a full static pass. Early-access partners, the post says, saw stronger results in those tests. The company published a short API sample that points Gemini 3.7 Flash at a YouTube URL and asks for the three most important announcements in a keynote.

Rohan Doshi, a senior product manager at Google DeepMind, and Mario Lucic, a research director there, signed the post. They credit Sergi Caelles, Filip Pavetic, Ahmet Iscen, Suhas Yogin, and the Agentic Vision team.

Decoded Take

Video is expensive in tokens, which is why most products still sample it thinly and miss the frame that matters. Google is selling a controller that decides which seconds to inspect, then charging the same list price. If the 88 percent token cut holds on customer footage, this becomes a default flag rather than a demo. The harder test is YouTube. Ask YouTube only works if answers stay tied to what is on screen, not to a hallucinated recap of the thumbnail. Watch whether the Gemini app ships the mode to free users, whether Enterprise customers leave “agentic” on for production pipelines, and whether Ask YouTube cites timestamps people can scrub to. A quality bump of 7 percent is easy to lose in marketing. A cheaper, timestamped answer on a two-hour keynote is not.