Gemini Adds Agentic Video Search With Fewer Tokens
Google’s Gemini models can now choose which moments of a video to inspect, reducing token use while improving retrieval accuracy.
The model decides where to look
Google DeepMind introduced agentic video understanding for Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Instead of sampling an entire recording at a fixed frame rate, the models first reason over available audio, transcript and visual context, then dynamically inspect the sections and frames most relevant to a question.
The process resembles tool use. Gemini can seek to particular timestamps, increase sampling around a suspected event and revisit the recording when its initial evidence is insufficient. Google says this selective approach improves accuracy while consuming up to 88% fewer tokens than static video processing. Lower token use can translate into lower cost and longer practical reach for applications examining meetings, surveillance archives, sports footage or instructional recordings.
The feature matters because fixed-rate sampling imposes an awkward compromise. Sampling densely captures brief events but makes long videos expensive; sampling sparsely saves tokens but may miss a single gesture, object or line of text. An adaptive system can devote computation to uncertain or information-rich moments, although its initial routing decision introduces another failure mode: a model that looks in the wrong place may never collect the evidence needed to correct itself.
Google is making the capability available through its latest Flash models and has published developer guidance for building agentic video-analysis workflows. The announcement concerns video understanding rather than generation, and it does not turn Gemini into an autonomous surveillance product by itself. Developers remain responsible for storage, permissions, tool access and any actions taken from the model’s conclusions.
Why it matters
Long-form video is one of the most expensive multimodal inputs because most frames add little value to a specific question. Allowing a model to control its own inspection schedule makes video analysis look more like information retrieval than exhaustive vision processing. If developers can measure when the search strategy misses decisive moments, the approach could make large video libraries economically searchable without preprocessing every frame.