On September 1, Google rolled out a video-processing mode called agentic video in the Gemini API, handing the question of "how to watch a video" back to the model itself. It's available to developers immediately, will roll out to the Gemini app soon, and YouTube's Ask YouTube feature is expected to adopt it in the coming months.
Three numbers
The change happens at the sampling layer. Until now, Gemini processed video at a fixed frame rate — one frame per second by default — feeding footage into the context window at the same pace whether it was a product launch or a security camera feed. Under the new mode, the model is equipped with native video tools and decides, as it reasons, which segments to look at, at what speed, and whether to pull visual frames, audio, or subtitles — retrieving only the moments it actually needs.
Google's benchmark results: token consumption down by as much as 88%, cost down by as much as 66%, and accuracy up by as much as 7%. The company also says Gemini 3.7 Flash now sits on the Pareto frontier for video analysis — the point where, at a given price, nothing is more accurate, and at a given accuracy, nothing is cheaper.
The third number is the most counterintuitive of the three. Downsampling usually means losing information — watch less, know less. Here it can push accuracy up instead, because fixed-frame-rate sampling wastes and misses at the same time: in a two-hour meeting recording, most frames show the same person in front of the same slide, and that repetition eats up the context budget, while the one moment that actually matters might last under a second — exactly the kind of thing 1 FPS sampling can miss. Pulling clips on demand cuts the redundant frames while letting the model sample more densely wherever it needs to.
A rough tally
An hour of footage at one frame per second works out to 3,600 frames; fed in one by one, the input token count starts in the six figures (by rough estimate). That number alone has kept a lot of use cases off the table — a three-hour livestream replay, a week of in-store camera footage, a full-length documentary — where just reading through the source material once already costs more than the task is worth. Cut that by 88% and it drops into the five figures, which is roughly the threshold where long-video analysis stops being a demo and starts being a product feature.
Google's repeated moves on the Flash line this year have mostly pointed at unit cost. Earlier progress on video was about pushing up how long a single clip could be processed; this is about pushing down what it costs per unit of length. Stack the two together and the effect is bigger than either move on its own.
Availability and how to turn it on
The feature works with Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, and is available through the Gemini API (via Google AI Studio) and the Gemini Enterprise Agent Platform. Billing still runs on standard Gemini API token pricing — there's no extra charge for the feature itself, so the token savings show up directly on the bill. The change on the developer side is minimal: just set the processing mode to agentic in the configuration.
The four use cases Google highlights point at what it's chasing: sub-second clip localization for editing; needle-in-a-haystack search across hours of footage; anomaly detection via dynamic resampling; and counting actions or objects accurately across a timeline. The first two lean toward content and media work; the latter two are clearly aimed at surveillance, quality inspection, and site-audit workloads.
What it means for video pipelines
Teams doing short-video distribution, livestream QA, or security-footage review have generally found the economics of long-video analysis didn't work — the standard workaround has been to run a cheap model or traditional computer vision as a first pass, then hand only the suspect clips to a large model. Once sampling decisions move to the model itself, that pipeline can lose a step: the first-pass filtering effectively gets folded into the model's own reasoning.
One caveat worth keeping: 88%, 66%, and 7% are all "up to" figures — best-case numbers under the most favorable conditions. Google hasn't published a distribution showing how much those gains shrink for different footage or task types. Anyone moving this into production should still run their own comparison on their own material first.
Sources: Google's official blog, CocoLoop, Google DeepMind. The 88% token reduction, 66% cost reduction, 7% accuracy gain, supported models, and billing details are all checked against the official announcement; the frame-count estimate is the editor's own rough calculation.