Gemini introduces 'Agentic Video Understanding,' a video analysis tool that automates video analysis, generating detailed and inexpensive answers by repeatedly transcribing and re-checking long videos.

On September 1, 2026, Google DeepMind announced 'Agentic Video Understanding,' a new feature that enables Gemini to autonomously analyze videos. Gemini searches for necessary scenes in a video based on the question, and constructs an answer by reviewing the video and checking the audio transcript. This feature is available in Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite.

Introducing agentic video understanding with Gemini

Can Gemini count the number of claps? 👏

Accurately counting rapid movements is a notoriously tricky task for AI. Because static processing ingests video at a fixed 1 FPS by default, split-second movements like a clap easily get missed entirely or get confused with a snap or… pic.twitter.com/qbv5o3wR7X

— Google AI Developers (@googleaidevs) September 1, 2026

Gemini has long included features that allow users to input videos, summarize their content, and ask questions about them. Traditional standard video processing used 'static processing,' which involved extracting one frame per second from the video and loading the video frames and audio together into Gemini. While this is easy to use for examining short videos, it tends to consume a large number of tokens for videos lasting several tens of minutes to several hours. On the other hand, reducing the amount of input information by lowering the frame rate may cause important information that only appears for a fleeting moment to be missed.

Agentic Video Understanding is a processing mode that extends existing video understanding capabilities, allowing Gemini itself to determine 'which parts of the video need to be examined.' Based on the question, Gemini searches for the necessary sections and extracts the required information from video frames, audio, and transcripts. If further examination is needed, it can adjust the frame rate and resolution of the target section and reread it. Instead of reading the video at the same density from beginning to end, it concentrates processing on the parts necessary for the answer.

Because it allows for detailed examination of only the necessary parts, it can be used for applications such as finding instantaneous changes. Google cites applications such as detecting state changes that occur in less than a second, identifying cut points in video editing, searching for specific events in several hours of footage, anomaly detection by increasing the frame rate only in sections where anomalies are found, and counting the number of actions or objects. When counting fast movements, the frame rate is changed as needed to re-examine the footage, allowing for detailed examination of movements that are difficult to capture with static processing at 1 FPS.

The image below compares token consumption and accuracy in three different video understanding benchmarks using Gemini 3.7 Flash with traditional static processing versus using Aggent Video Understanding. According to Google, Aggent Video Understanding can reduce token consumption by up to 88% and improve accuracy by up to approximately 7%. The efficiency improvement is particularly significant for relatively long videos, such as explanatory videos of about 10 minutes, lectures of 90 minutes, and recordings lasting several hours.

The image below shows a graph comparing the cost and accuracy per query in '1H-VideoQA,' which measures the ability to understand long-duration videos. The higher up the graph is, the lower the cost and the higher the accuracy. Google explains that Gemini 3.7 Flash with Aggent Video Understanding enabled is at the 'Pareto front line' in terms of accuracy and cost among the models they have tested. According to Google, Aggent Video Understanding can reduce analysis costs by up to 66%.

To use Agentic Video Understanding, simply specify 'agentic' as the video processing method in the Gemini API. As of the time of writing, it is available in Google AI Studio and Gemini Enterprise Agent Platform, and supports the analysis of uploaded videos as well as public YouTube videos. There is no additional charge for Agentic Video Understanding itself; the usual Gemini API token fees apply. Note that traditional static processing is also available for short videos where response speed is important, and Google's developer documentation suggests that static processing is an option for short videos under 5 minutes.

Furthermore, Google plans to extend Aggent Video Understanding beyond just a developer API to include Flash and Flash-Lite versions of the Gemini app in the near future. Within the next few months, it will also be implemented in 'Ask YouTube,' a feature allowing users to ask questions on YouTube video pages, to improve the quality of answers based on the video content.