Google gave Gemini a video tool that picks what to watch, and reports up to 88% fewer tokens

Until September 1, 2026 a video entered a Gemini model at a fixed frame rate, one frame per second by default and adjustable through the API. Every second of a two-hour recording was paid for in tokens whether anything happened in it or not. Agentic video understanding replaces that with a loop: the model calls an internal tool to load the part of the file it needs, sets the frame rate for that part, and chooses whether to look at frames, listen to the audio, or read the transcript.

What changes between fixed sampling and the agentic loop

Static processing takes in the whole stream at one rate, so a developer chooses between a large token bill and a sampling shortcut that drops the moment that mattered. Google says the agentic path removes that choice, because the model searches, scans, and inspects target segments itself. The post also notes that a developer could already do this by hand; what is new is that the model runs the loop, which removes the orchestration code around it.

The savings are largest on long material, which Google describes as everything from a 10 minute how-to guide through a 90 minute lecture to multi-hour recordings. One possible explanation is that a short clip already fits inside a fixed sampling budget, so there is less to skip.

• Sub-second moment retrieval, for state changes and cut boundaries a one-frame-per-second pass walks past.
• Needle-in-a-haystack questions over multi-hour video without spending millions of tokens.
• Anomaly detection, by resampling an interesting window at a higher frame rate.
• Counting repeated movements and distinct objects across time.

The numbers Google published, and the ones it did not

On standard video analysis benchmarks, Google reports up to 88% lower token consumption, up to 66% lower cost, and up to 7% better accuracy. All three are the vendor's own measurement, and all three are ceilings rather than averages.

The post names one benchmark, LongVideoBench, in a demonstration of Gemini 3.7 Flash with the feature on and off. It publishes no table of benchmarks with per-benchmark values, and it does not say which baseline frame rate the comparison ran against. Anyone budgeting from the 88% figure is budgeting from a number whose worst case is not published.

Google also places Gemini 3.7 Flash with agentic understanding on the accuracy-to-cost pareto frontier among the models it tested for video understanding, and calls it the best of the three supported models on quality. Which models were in that comparison is not stated.

How to turn it on, and where it goes next

The feature has been on the Gemini API since September 1, 2026, in Google AI Studio and in the Gemini Enterprise Agent Platform, for uploaded video files and for YouTube videos. A developer enables it by setting the processing field on a video input to agentic, and the rest of the request stays as it was.

Billing runs at standard Gemini API token rates with no separate charge for the feature, so the saving arrives as a smaller token count rather than a lower price per token. Google says the same capability will reach Gemini app users on the Flash and Flash-Lite models soon, and will power YouTube's Ask YouTube answers on the watch page in the coming months.