Google introduces a new way to analyze videos with Gemini. The feature, called agentic video understanding, allows the model to decide which segments to review, how quickly to analyze them, and whether it needs to use images, audio, or transcripts.
The new capability is coming to Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite through the Gemini API, Google AI Studio, and Gemini Enterprise Agent Platform. The main promise? Analyze long videos with greater precision, fewer tokens, and lower costs.
From static analysis to active understanding
Until now, a common way to process a video was to extract frames at a fixed rate. The API's default setting is 1 frame per second, although developers can modify it.
This approach works for general questions, but it can miss important details. A change lasting only a few moments, an object that appears briefly, or a visual anomaly could fall between two analyzed frames.
With agentic understanding, Gemini takes on a more active role. Instead of reviewing the entire video uniformly, it uses internal tools to search, scan, and inspect the segments related to the question.
Gemini does not watch every second of the video in the same way. It decides where to focus its attention and which signal it needs to consult.
The model can combine frames, audio, and transcripts within a reasoning loop. This architecture resembles the agentic vision approach, which combines native image understanding with code execution to solve more complex visual tasks.
Fewer tokens and greater precision
According to tests shared by Google, agentic video understanding can reduce analysis costs by up to 66% and token consumption by up to 88%. At the same time, the company reports precision improvements of up to 7%.
These results are especially relevant for long videos, such as 10-minute tutorials, 90-minute classes, meeting recordings, or files lasting several hours. In these cases, processing every frame can be too expensive, while analyzing only a few frames increases the risk of missing information.
Google says that Gemini 3.7 Flash offers the best combination of quality and cost among the models evaluated. In technical terms, the company places it at the efficiency frontier between accuracy and price—in other words, in a strong position for applications that need good results without sending inference costs through the roof.
What tasks can it handle?
The new capability is designed for situations where finding specific details within a large amount of audiovisual material matters.
- Retrieving moments lasting less than a second: It can locate rapid state changes or precise cuts, which is useful for automatic video editing.
- Complex searches in long videos: It can answer specific questions about recordings lasting several hours without consuming millions of tokens.
- Anomaly detection: It can resample a time window at a higher frame rate to review rapid movements or visual artifacts.
- Counting actions and objects: It makes it easier to track repeated movements and count distinct objects over time.
Imagine, for example, a company that wants to know the exact moment when a failure appears on a production line recording. With static analysis, it would have to increase the number of frames across the entire video or accept that it might miss the problem. With the agentic approach, Gemini can identify a suspicious area and review it in greater detail.
How to activate it in the Gemini API
The feature does not require an additional fee. Google says it uses the Gemini API's standard token pricing. To activate it, the developer must set the processing mode to agentic in the video configuration.
An official Python example uses a YouTube video and asks a question about its main announcements:
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-3.7-flash",
input=[
{
"type": "video",
"uri": "https://youtu.be/7Z5Vy9JBANs",
"processing": "agentic"
},
{
"type": "text",
"text": "What are the 3 most important announcements in this keynote?",
},
],
)
print(interaction.output_text)
The same capability is available for both user-uploaded videos and YouTube videos, always through Google's compatible platforms and services.
It will also come to Google's products
Google says it will bring these efficiency and quality improvements to billions of users. The feature will soon roll out in the Gemini app for Flash and Flash-Lite models.
In addition, over the coming months, the technology will begin powering Ask YouTube, a feature that will let users ask questions from the playback page. The answers will aim to make better use of the video's visual elements, rather than relying only on its audio or transcript.
The change shows where multimodal analysis is heading. AI is no longer limited to receiving a file and processing it from beginning to end. It can establish a search strategy, decide what evidence it needs, and focus its resources on the relevant moments.
For developers, that means less custom code is needed to build sampling and search systems. For users, it could translate into more precise answers about long videos. The important question is no longer whether an AI can watch a video, but whether it knows where to look and why.