On September 1, 2026, Google DeepMind announced "Agentic Video Understanding," a new feature designed to allow Gemini to autonomously carry out video analysis.
This feature is available for Gemini 3.7 Flash, Gemini 3.6 Flash, and Gemini 3.5 Flash-Lite.
Conventional video processing typically relied on "static processing," where one frame is extracted per second of video. However, this approach faced challenges with long-form videos spanning several hours, as it consumed a massive number of tokens. Conversely, lowering the frame rate risked missing critical, fleeting information.
Agentic Video Understanding is a mechanism that allows Gemini to determine for itself "where in the video it should look." Based on a query, the model searches for the necessary segments and retrieves information from video frames, audio, and transcripts. If necessary, it can re-read the target segments by adjusting the frame rate or resolution.
According to Google, this method reduced token consumption by up to 88% and improved accuracy by up to approximately 7%. The company further explained that analysis costs could be reduced by up to 66%. The improvements in efficiency are reported to be particularly significant for long-form content, such as 10-minute instructional videos or recordings spanning several hours.
To use this feature, the video processing method must be specified as "agentic" via the Gemini API. It is currently available in Google AI Studio and the Gemini Enterprise Agent Platform, and it also supports the analysis of public YouTube videos. Pricing follows the standard Gemini API token rates.
Google plans to roll out this feature to the Flash and Flash-Lite versions of the Gemini app in the future. Additionally, the company stated that it will be integrated into "Ask YouTube," which allows users to ask questions on YouTube video watch pages, within a few months.
Source: Geminiが動画を自律的に分析する「Agentic Video Understanding」登場、文字起こしや再確認を繰り返して長時間動画を詳しく分析しつつ安価に回答を生成 (GIGAZINE, 2026-09-02)