via MarkTechPost
Google Launches Agentic Video Understanding for Gemini Flash Models, Slashing Video Tokens by Up to 88%
Video has long been the most costly modality for AI models to reason over. Previously, when a Gemini model was given a 90-minute lecture, it would process the entire video at a fixed rate of one frame per second, regardless of the task—whether summarizing the content or pinpointing when the speaker transitions to a pricing slide. This single-pass approach forced a compromise: either pay for the full video timeline in context or pre-chunk the video and risk losing critical details.
This week, Google launched agentic video understanding for its Flash models, marking a shift in how Gemini handles video. Instead of ingesting the entire timeline, Gemini now navigates it, deciding what to watch, at what frame rate, and through which modality. According to Google, this approach delivers up to 88% reduction in tokens, up to 66% lower cost, and up to 7% higher accuracy on standard video benchmarks.
Is it deployable? Yes, but exclusively as a hosted API feature—there are no open weights or self-hosting options. It is available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, supporting both file uploads and public YouTube URLs. Pricing follows standard Gemini API token rates with no additional feature fees.
What Actually Changed
Static video processing, which remains the default across Gemini models, extracts frames at 1 frame per second (FPS) in a single pass, processes audio at 1 kilobit per second (Kbps) in mono, and inserts timestamps every second. Agentic processing replaces this with a dynamic loop. The model combines its reasoning with native video tools to selectively analyze content. This allows Gemini to focus on relevant segments, adjust frame rates based on visual complexity, and even switch between visual and audio modalities as needed.
The result is a more efficient and intelligent approach that not only reduces resource consumption but also improves performance metrics across the board. For developers and enterprises dealing with extensive video datasets, this represents a significant upgrade in both cost-effectiveness and analytical precision.
As AI video understanding evolves, particularly into 2026, the ability to process long-form content without exhaustive computational overhead becomes increasingly critical. Google's agentic approach positions Flash models as a viable option for real-time video applications, from automated content moderation to contextual search in media archives. With over 200 million developers already leveraging Gemini APIs, the impact of this feature could be substantial.
For technical teams, the immediate next step is to experiment with the agentic mode via the API to benchmark its performance against traditional methods on their specific use cases. The reported accuracy gains—up to 7% better on standard benchmarks—suggest that selective processing does not compromise understanding, but rather enhances it by allowing the model to allocate resources where they matter most.