Google launches agentic video understanding for Gemini, cutting token usage by up to 88%

Google DeepMind introduced a capability called agentic video understanding for its Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. [2] The feature utilizes the native video tools of the models to dynamically inspect, search, and scan target video segments across audio, transcripts, and visual frames. [2] This dynamic processing replaces previous static methods that ingested video at a fixed rate, typically one frame per second. [2]

Across standard benchmarks, the agentic approach reduces token consumption by up to 88 percent and analysis costs by up to 66 percent. [2] Additionally, Google reports that the capability improves accuracy by up to 7 percent. [2] The feature is currently available for YouTube videos and video uploads through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. [2] Google DeepMind also describes the broader Gemini 3.5 family as models combining frontier intelligence with action. [1]

By shifting from static frame extraction to dynamic scanning, Google is directly addressing the high token overhead typically associated with multimodal video analysis. This approach allows developers to process longer video contexts without linearly scaling their API costs, making the Gemini Flash tier highly competitive for video-heavy enterprise workloads. The integration into both Google AI Studio and the Enterprise Agent Platform suggests a push toward making advanced tasks, like sub-second moment retrieval and anomaly detection, more commercially viable. What the sources don't address: Whether the dynamic searching and scanning processes introduce additional latency compared to traditional fixed-rate ingestion methods.