Google has launched agentic video understanding for Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite. The feature lets a model search, scan, and inspect selected portions of a video instead of processing every segment at a fixed rate. In its September 1 announcement, Google says the approach can cut token consumption by up to 88%, reduce analysis costs by up to 66%, and improve accuracy by up to 7% on its standard video-analysis benchmarks.
The capability is available now for video uploads and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Google’s developer documentation shows the control as an agentic processing setting. That makes this a change to how developers ask Gemini to analyze video, rather than a new model tier or a separate fee.
From fixed sampling to targeted inspection
Traditional video analysis samples a recording at a configured frame rate. Google describes the default as one frame per second, with the rate adjustable through the API. That method is straightforward, but long recordings can create a poor choice: send enough frames to catch a brief event and pay for a large context, or sample more lightly and risk missing the detail that answers the question.
Agentic video understanding gives Gemini native video tools that it can use during reasoning. The model can search for likely moments, inspect visual frames, draw on audio, and use transcripts to focus its work. Google points to sub-second moment retrieval, anomaly detection, and precise counting as examples of the tasks this approach is meant to improve. The company’s published benchmarks are its own results, so the percentages describe Google’s tested configurations rather than a universal performance guarantee for every video or prompt.
The long-form use case is the clearest target. Google says the efficiency gains are especially pronounced for recordings ranging from 10-minute how-to guides to 90-minute lectures and multi-hour material. A developer could ask for the important announcements in a keynote, locate every appearance of a component in a training video, or find a spoken detail without first building a separate indexing pipeline.
A smaller bill for video-heavy applications
Google says Gemini 3.7 Flash with agentic understanding offers the strongest quality and cost combination among the tested models. The feature uses standard Gemini API token pricing with no additional feature fee, according to the announcement. The actual bill still depends on the model, input, output, and the work performed, while the claimed savings depend on the workload and comparison method.
The rollout also extends beyond developer tooling. Google says agentic video understanding will reach the Gemini app across Flash and Flash-Lite models soon. In the coming months, it plans to use the capability in YouTube’s Ask YouTube feature on the video watch page, where answers can be grounded in the visual content as well as other video signals.
That puts the update alongside Google’s broader push to make Gemini a tool-using system. Our earlier coverage of Gemini 3.7 Flash’s launch focused on coding and agent workflows, while our report on Antigravity’s Gemini agent teams examined collaborative software agents. This video release applies the same general idea to a medium where the useful answer may exist for only a few seconds.
For now, the practical test is simple: developers can set video processing to agentic and compare cost, latency, and answer quality against their existing workflow. Google has published the interface and benchmark claims, but production teams will still need to measure their own footage, prompts, and accuracy requirements.
