Google introduces agentic video understanding for Gemini flash models, reducing token usage by 88%

Google has introduced agentic video understanding for its Gemini Flash models, significantly reducing video token usage by up to 88%. This new feature allows the models to navigate video content dynamically, deciding what to watch, at what frame rate, and through which modality, instead of processing the entire timeline at a fixed rate. The change results in up to 66% lower costs and 7% higher accuracy on standard video benchmarks. The agentic processing replaces static processing, which extracts frames at 1 FPS and processes audio at 1 Kbps. Instead, the model pairs its own reasoning with native video tools to search, scan, and inspect target segments across frames, audio, and transcripts, loading only what the prompt requires. This approach enables developers to mix modes per video within a single request, using agentic processing for long lectures and static processing for short clips. The feature is available via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, supporting both file uploads and public YouTube URLs. Token accounting splits into thought tokens and tool-use tokens, with no additional fees for the agentic feature. Google highlights that Gemini 3.7 Flash with agentic understanding achieves a balance of accuracy and cost efficiency for video analysis.