Google launches agentic video understanding with Gemini

Google has enabled a new way to analyze videos with Gemini that can reduce token usage by up to 88%, lower costs by up to 66%, and improve accuracy by up to 7%, according to the company's tests.

The feature is called agentic video understanding and is now available to developers through the Gemini API, Google AI Studio, and Gemini Enterprise Agent Platform. It works with Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite.

Gemini stops reviewing video at a fixed rate

Until now, video analysis worked much like flipping through pages at regular intervals. The model received images at a fixed rate, usually one frame per second, although developers could change that value.

That approach can miss important details. A gesture lasting less than a second, a quick change on a screen, or an object that appears for an instant may not be recorded.

With the new system, Gemini decides which parts it needs to review and how to review them. It can search for a specific moment, increase the sampling rate within an interval, check the audio, or use the transcript, depending on the question it needs to answer.

The difference is similar to searching for a specific scene in a movie instead of looking at one image every second throughout the entire playback. The model analyzes only the relevant parts instead of processing the whole video with the same level of intensity.

What it can do

Google is positioning this capability for tasks that are often expensive or inaccurate, especially in long videos such as courses, meeting recordings, tutorials, or conferences lasting several hours.

• Find moments lasting less than a second, which is useful for locating exact changes and automating video cuts.
• Answer questions about lengthy videos, even when the information is spread across several hours.
• Detect anomalies by reviewing scenes with fast movements or unusual visual details at a higher speed.
• Count objects and actions, such as exercise repetitions, products on a production line, or specific movements.

In a conference video, for example, you could ask what the three main announcements were. In a factory recording, the system could search for when unusual behavior appears without having to study every second at the same level of detail.

Lower costs for long videos

The results published by Google come from benchmark tests for video analysis, not a guarantee for every file or query. In those tests, the system reduced token usage by up to 88%, analysis costs by up to 66%, and improved accuracy by up to 7%.

The improvements were clearer in long videos. With traditional processing, developers had to choose between paying more to preserve details or reducing the number of images analyzed and risking the loss of information.

Google says Gemini 3.7 Flash offers the best combination of quality and cost among the models evaluated. The feature uses the API's standard pricing and does not add a specific fee.

Now available to developers

To enable it, developers must specify agentic in the API processing configuration. It can be used with both uploaded videos and YouTube videos.

This reduces the technical work required. Previously, a team might have had to program interval searches, changes in analysis speed, and the selection between images, audio, and transcripts on its own. Now Gemini can coordinate that process internally.

For you, the change still depends on where you use Gemini. The feature will soon reach all users of the Gemini app in the Flash and Flash-Lite models. Google is also preparing to bring it to the "Ask YouTube" feature, which will answer questions about a video using not only its transcript but also what appears on screen.

The important step is not that Gemini can watch videos. It could already do that. The novelty is that it can now decide where to look and which signals it needs, which could make analyzing hours of content more practical and affordable. What remains to be seen is whether that accuracy improvement holds up outside Google's tests and with real videos containing blurry images, poor audio, or ambiguous questions.