Google has unveiled a new feature called agentic video understanding for its Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. This innovation allows the models to dynamically analyze video segments, significantly enhancing accuracy while reducing token usage by up to 88% and costs by up to 66%. Users can activate this feature by setting their API configuration to “agentic” in Google AI Studio or the Gemini Enterprise Agent Platform.
The agentic video understanding feature uses Gemini’s built-in video tools to provide improved performance and unlock new video processing capabilities. These include the ability to retrieve specific moments quickly, detect anomalies with greater accuracy, and perform precise counting tasks. The feature is available for both video uploads and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
Instead of the traditional static processing method, which ingests video at a fixed frames-per-second rate, agentic video understanding leverages the model’s core reasoning abilities to scan and inspect video segments dynamically. This approach pairs visual frames, audio, and transcripts, enabling a more efficient and accurate analysis. Across video analysis benchmarks, this method reduces costs by up to 66% and token consumption by up to 88%, while also improving accuracy by up to 7%.
These improvements are particularly noticeable in long-form videos, ranging from brief how-to guides to lengthy lectures and recordings. The static method typically forces developers to manage high token costs or resort to strategies that might overlook important details. However, by activating agentic video understanding, token consumption drops significantly, and accuracy sees an increase of up to 7% with Gemini 3.7 Flash.
Agentic video understanding allows the model to actively determine which parts of a video to focus on, adjusting the speed and modality (such as frames, audio, or transcript) as needed. This capability not only reduces development overheads but also transforms how long-form video content is processed across various applications.
For instance, when assessing the LongVideoBench benchmark, Gemini 3.7 Flash demonstrates considerable token reduction and accuracy improvements with agentic video understanding activated. This model can also accurately analyze fast-paced actions by adjusting the frames-per-second dynamically.
The feature is available through the Gemini API and does not incur additional fees beyond standard API token pricing. Developers can enable this by setting processing to “agentic” in their API configurations. Google plans to extend the efficiency and quality benefits of this feature to billions of users across its products, with upcoming integration into the Gemini app and YouTube’s ‘Ask YouTube’ feature.
Many early access partners have reported strong performance results using agentic video understanding. The team behind this development includes Sergi Caelles, Filip Pavetić, Ahmet Iscen, Suhas Yogin, and the Agentic Vision team.