Google adds an 'agentic' mode to Gemini video analysis, claiming up to 88% fewer tokens

Google rolled out an agentic mode for video analysis in Gemini on September 1, letting the model decide for itself which parts of a video are worth inspecting instead of grinding through every frame at a fixed rate. Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite can now request specific frames, audio segments, or transcript chunks on demand -- an internal loop the model runs before answering a query about a video, rather than ingesting the whole thing uniformly.

Google frames four jobs the agentic approach targets: catching sub-second state changes for precise video editing, answering complex queries across multi-hour footage without processing all of it at one frame rate, resampling suspect windows at higher frame rates to catch visual anomalies, and counting repeated actions or objects over time. The efficiency numbers -- up to 88% fewer tokens consumed, up to 66% lower cost, up to 7% better accuracy -- are Google's own reported benchmarks; independent verification hasn't caught up yet, and outside coverage of the launch has so far only restated the company's own figures rather than testing them.

The bigger test of the claim is what happens when it reaches products people use daily rather than developers calling an API. Google says the mode will roll out to the consumer Gemini app "soon" and will power YouTube's "Ask YouTube" answer feature on video watch pages "in the coming months" -- though YouTube's own help documentation still doesn't explain how a video gets chosen as Ask YouTube's primary citation versus a supporting one, leaving creators an efficiency gain they can't see yet and no clear way to influence how their videos get read.