Google rolled out agentic video understanding across three models on Tuesday: Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Google puts the savings at up to 66 percent on the cost of analysing a video and as much as 88 percent on tokens burned, with accuracy up 7 percent at best. All three figures are its own. It is live now through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
The mechanism is a change in how the model samples a video, not a new model architecture. Standard Gemini processing ingests footage at a fixed rate, one frame per second by default, regardless of what is happening on screen. Agentic video understanding instead lets the model call an internal tool to jump around a clip: it decides which segments to inspect, at what frame rate, and whether to check the visual track, the audio, or the transcript, according to Google.
That distinction matters most on long footage. The savings concentrate, Google says, in long material: a 10-minute tutorial, a lecture running an hour and a half, a recording that goes on for hours. Those are exactly the cases where sampling at a fixed rate makes a developer pick between an enormous token bill and missing something. Among the three models, Google positions Gemini 3.7 Flash as the strongest tradeoff of accuracy against cost, calling it the frontier option on that axis in its own testing.
Google lists four target use cases: sub-second moment retrieval for precise video editing, needle-in-a-haystack search across multi-hour footage, anomaly detection through variable frame-rate resampling of suspect windows, and counting repeated actions or objects over time.
Counting and anomaly detection over long video are the two jobs enterprises have been paying specialist vendors to handle: manufacturing-line defect spotters, retail loss-prevention tools, sports and traffic analytics built on custom computer-vision pipelines. A general-purpose model that does both adequately, inside a standard API call with no added fee, compresses a category that used to justify a dedicated vendor relationship. The caveat is that Google’s own numbers describe averages across its test set, and “adequate” on a benchmark is not the same as reliable enough to flag a safety defect or a shrinkage event without human review.
That gap is worth naming directly. Every figure in this launch, the 66 percent cost reduction, the 88 percent token drop, the 7 percent accuracy gain, comes from Google’s internal evaluation. The announcement does not include third-party or independently audited benchmark results, and Google has not disclosed which datasets beyond LongVideoBench made up the comparison set.
Activation is a one-line change: developers set the processing field to “agentic” in the API call, and pricing follows standard Gemini token rates with no separate fee. Everyone using the Gemini app on the Flash and Flash-Lite tiers gets it soon, Google says, and it will sit behind the “Ask YouTube” box beside a video within months.
Teams currently running video analysis on fixed-rate Gemini calls, or paying a specialist vendor for counting or anomaly detection on long footage, should benchmark agentic processing against their own error tolerance before treating Google’s accuracy numbers as a substitute for that vendor relationship.
Google detailed the launch in a September 1, 2026 post on its own blog.
