Google announced the launch of “Agentic Video Understanding” across Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite, making it immediately available for analyzing uploaded files and YouTube clips through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. The company says the feature can reduce token consumption by up to 88%, cut analysis costs by up to 66%, and improve accuracy by up to 7% in benchmark tests.
The update targets a fundamental problem in long-video analysis: Traditional processing handles a clip at a fixed frame rate, defaulting to one frame per second, which can be adjusted through the API. This approach may miss fast moments or details that do not appear in the selected sample, while increasing the frame rate raises the number of tokens and the cost.
How does the agentic mode work?
Instead of ingesting the entire video stream in a fixed manner, Gemini can determine what should be watched, the appropriate inspection speed, and the most useful medium for the task: frames, audio, or transcribed text. The model calls an internal tool to load relevant portions of the video file within a goal-directed loop, then reexamines specific time windows when necessary.
Google explains that developers were able to build some of these mechanisms manually, but moving the dynamic search and inspection process into the model reduces the required programming effort. The feature does not impose an additional fee; it uses standard Gemini API token prices, and it is enough to set the processing option to agentic in the API settings.
What changes for developers in practice?
Google says the benefits are most apparent with long-form content, from 10-minute instructional guides and 90-minute lectures to recordings lasting several hours. The uses highlighted by the company include:
- Retrieving moments more precisely than a second, including state changes and rapid cut boundaries needed for automated editing.
- Searching for specific information within a long video or a multi-hour recording without consuming millions of tokens.
- Detecting anomalies by resampling important time windows at a higher frame rate.
- Counting repeated movements and objects more accurately by tracking them over time.
According to Google’s tests, Gemini 3.7 Flash with agentic understanding delivers the best overall quality and the best quality-cost combination among the tested models, placing it on what the company describes as the efficiency frontier between accuracy and cost. These results remain figures reported by the company and are not a guarantee of identical performance for every type of video or task.
Why does this news matter?
The actual change is not the addition of a new model, but the transformation of video analysis from fixed, comprehensive processing into adaptive selection of sections and signals. This could benefit applications that rely on searching long archives or monitoring fast events, because the reduction in tokens does not come from ignoring the video altogether, but from directing the inspection toward the clips the model considers relevant to the question.
At the same time, developers still need to assess the accuracy of agentic selection in scenarios where the important moment is rare or unclear, and the source does not provide independent details about the full testing methodology or absolute cost values. Therefore, the improvement percentages should be read as benchmark results reported by Google, not as a substitute for testing the feature on each application’s content.
Availability and expansion plans
The feature is currently available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform for the three models mentioned. Google plans to roll out efficiency and quality improvements to all Gemini app users soon through Flash and Flash-Lite models, and will also use them in the coming months in the “Ask YouTube” feature on the video watch page, with the aim of providing higher-quality answers based on the visual content in the clip.