Google Adds Agentic Video Understanding to Gemini, Claiming 88% Lower Token Use

Google has launched agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. The feature lets the models search selected parts of a video, using images, audio and transcripts instead of processing the full recording at a fixed frame rate. Google says this can cut token use by up to 88% on standard video-analysis benchmarks.

The capability is available now for uploaded videos and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Developers enable it by setting video processing to “agentic”; Google says it uses standard Gemini API token pricing and carries no separate feature fee.

From uniform sampling to targeted inspection

Static video processing ingests media at a set frames-per-second rate. Google says agentic video understanding instead combines Gemini’s reasoning with native video tools, allowing the model to decide what moment to load, how quickly to inspect it, and whether frames, audio, or a transcript offer the relevant signal.

The change affects both workflow and model behavior. Developers could previously build targeted inspection themselves, Google says. The new mode moves that selection into an internal agentic loop for long videos, where fixed-rate processing can trade higher token use for missed details.

Up to 88% lowerToken consumption

On standard video-analysis benchmarks, Google reports up to 88% lower token consumption.

Up to 66% lowerAnalysis cost

The company also reports analysis costs up to 66% lower on those benchmarks.

Up to 7% higherAccuracy

Accuracy was up to 7% higher, according to Google’s benchmark results.

The jobs that benefit from a second look

Google identifies four intended uses: retrieving a precise moment in under a second, searching long recordings for hard-to-find details, detecting anomalies, and counting actions or objects. Each relies on selective resampling. The system can revisit a short window at a higher rate when fast motion or a subtle change requires closer inspection.

• Sub-second moment retrieval is intended to find brief state changes and cut boundaries that a one-frame-per-second pass can miss.
• Long-form search is intended to answer complex questions across multi-hour video without processing the recording uniformly.
• Anomaly detection and action or object counting can use closer inspection of relevant time windows.

A configuration choice, not a new price tier

The product claim is not that every video needs deeper analysis. It is that the model can fetch only the moments and modalities needed for a particular request. Google’s cost, token, and accuracy figures are company-reported benchmark results, so they describe tested performance rather than a guarantee for every video or task.

A route from developer API to consumer surfaces

Google plans to roll the feature out to Gemini app users on Flash and Flash-Lite models. It also says agentic video understanding will power YouTube’s Ask YouTube feature on video watch pages in the coming months, with answers intended to be grounded in video visuals.