Google added a new feature for video analysis in Gemini

Google has launched agentic video understanding — a new feature for the Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. It reduces token consumption for video analysis by up to 88%, processing costs by up to 66%, and increases accuracy by up to 7%.

Standard video processing operates at a fixed frame rate — one frame per second by default, with the parameter adjustable via the API. Agentic video understanding works differently: the model itself decides which parts of the video to watch, at what speed, and through which channel — frames, audio, or transcript — requesting only the necessary moments. The feature is similar to agentic image recognition, which Google released earlier and which combines code execution with native image understanding in Gemini models.

The greatest impact of agentic video understanding is observed on long videos — from ten-minute instructions to one-and-a-half-hour lectures and multi-hour recordings. With standard processing of such videos, developers have to choose between high token costs and loss of important details.

Among the new capabilities are precise search for specific moments with accuracy down to fractions of a second, finding answers to complex queries in multi-hour videos without consuming millions of tokens, anomaly detection through re-sampling interesting segments at higher frame rates, and accurate counting of repeated actions and objects.

The feature is already available for uploaded videos and YouTube clips via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Pricing has not changed — standard Gemini API prices apply, with no separate fee for the feature. To enable it, set the processing value to agentic in the API settings.

In the near future, agentic video understanding will become available to all users of the Gemini app for the Flash and Flash-Lite model series, and will also be used in the Ask YouTube feature on the video watch page.

Source: Google