Google has made the “agentic video understanding” feature available for the Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. Instead of processing the video from start to finish at a fixed frame rate, the system dynamically searches across visual frames, audio, and transcripts, examining only the necessary sections. This method reduces token consumption by up to 88% and costs by up to 66% while improving analysis accuracy, delivering up to a 7% increase in accuracy.
The feature offers capabilities such as finding sub-second moments, more accurate anomaly detection, and precise counting. It is particularly aimed at improving efficiency with long videos, ranging from 10-minute guides to 90-minute lessons and recordings lasting several hours. Among the supported models, Gemini 3.7 Flash delivers the best results in terms of overall quality and the quality-cost balance.
In the current static processing method, the model scans the video by default at 1 frame per second; this rate can be changed through the API. In the agentic approach, the model decides for itself what to watch, at what speed, and through which channel, loading the relevant video section with an internal tool. This reduces the need for developers to set up these processes manually. The system can rescan fast-moving sequences at different frame rates and find specific information in videos using fewer tokens.
The feature is available for video uploads and YouTube videos through the Gemini API on Google AI Studio and Gemini Enterprise Agent Platform. To enable it, the processing type must be set to “agentic” in the API configuration; standard Gemini API token pricing applies, and there is no additional feature fee.
Google will soon make efficiency and quality improvements available to all users of the Flash and Flash-Lite models in the Gemini app. In the coming months, the technology will also be integrated into the “Ask YouTube” feature on YouTube’s watch page.
Why it matters
This change reduces the technical burden of the analysis process for developers working with long video archives; the model’s ability to determine which section and data channel to examine limits the need to build custom scanning logic. Lower token usage and costs could make it easier to conduct detailed searches, counts and anomaly investigations on recordings lasting several hours within the existing API pricing. However, developers must separately select the processing type as agentic in the API configuration to use the feature; as a result, the standard video processing workflow does not change automatically. For now, access appears to be limited to Google AI Studio and Gemini Enterprise Agent Platform through the API, while it remains unclear when the announced rollout for use in the Gemini app and on YouTube will be completed.
Term: frame rate
Frame rate is the number of images drawn per second; as it increases, motion appears smoother.