Agentic Video Understanding in Gemini uses an active reasoning loop to dynamically scan video segments, reducing token consumption by up to 88% while improving accuracy for long-form content.
What the product does and how it is positioned
Agentic Video Understanding in Gemini improves the efficiency and accuracy of video analysis by moving away from fixed-rate frame ingestion. The model takes an active role in determining which specific moments and signals are required to answer a query.
This feature is integrated into the Gemini Flash model family and is accessible through developer platforms. It is optimized for long-form content, such as lectures and multi-hour recordings, where traditional static processing often results in high token costs.
Source-supported ways to use the product
Pinpointing split-second state changes and tight cut boundaries that are typically missed at standard frame rates.
Detecting anomalies by resample-ing specific time windows at higher frames-per-second to inspect rapid motion.
Accurately counting repeated movements or tracking distinct objects over the course of a video.
The documented workflow, where available
The developer sets the video processing parameter to 'agentic' within the Gemini API configuration settings.
The user provides a video source, such as a file upload or a YouTube URL, along with a natural language prompt.
The model uses an internal tool to fetch only the relevant segments of the video, audio, or transcript needed for the task.
The system returns the requested information, such as a summary, a specific timestamp, or a count of events.
Traditional video analysis models typically use static processing, where the video is ingested at a fixed rate. This approach can be inefficient for long videos, as it either consumes a massive number of tokens or misses critical details that occur between the sampled frames.
Agentic video understanding changes this by allowing the model to act as an agent that decides what to watch and at what speed. By invoking internal tools to load only the necessary data, the model can achieve higher accuracy with significantly lower resource consumption.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
The feature is available for Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite models.
It reduces costs by up to 66% by fetching only the necessary video segments rather than processing the entire stream at a fixed rate.
Yes, the feature supports both direct video uploads and YouTube videos via the Gemini API and Google AI Studio.
The model can dynamically choose to inspect visual frames, audio tracks, or video transcripts depending on the goal.
No, the feature uses standard Gemini API token pricing with no additional feature-specific fees.
