At the moment, we’re launching agentic video understanding throughout our newest fashions: Gemini 3.7 Flash, 3.6 Flash and three.5 Flash-Lite. This new functionality improves accuracy whereas dramatically lowering token utilization and prices for video evaluation. Just like agentic imaginative and prescient, which mixes code execution with Gemini fashions’ native picture understanding, agentic video understanding makes use of Gemini’s native video instruments to enhance efficiency and unlock new capabilities for video processing like sub-second second retrieval, extra correct anomaly detection, exact counting and extra.
The characteristic is offered at the moment for video uploads and YouTube movies through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
Benchmarks
In contrast to present ‘static’ processing, the place the mannequin ingests the video at a set frames-per-second price (default 1 FPS, adjustable through API), agentic video understanding pairs the mannequin’s core reasoning with native video instruments to dynamically search, scan, and examine goal video segments throughout visible frames, audio, and transcripts. Throughout customary video evaluation benchmarks, Gemini fashions with agentic video understanding cut back evaluation prices by as much as 66% and token consumption by as much as 88%, whereas bettering accuracy by as much as 7%.
These effectivity features are particularly pronounced on long-form video (from 10-minute how-to guides to 90-minute lectures and multi-hour recordings), the place static processing forces builders to decide on between excessive token prices or strategies that drop essential particulars.
