- Published on
Google Gemini's new agent-based video analysis cuts token usage by up to 88 percent
- Authors
- Name
- DeepHorus
Google has introduced a significant advancement in video analysis for its Gemini Flash models, employing an agent-based approach that dynamically searches through video footage. This new method, which replaces the previous fixed frame-by-frame scanning, is claimed to reduce token usage by up to 88 percent, cut associated costs by 66 percent, and simultaneously improve analysis accuracy. The feature is designed to make processing long-form video content both more efficient and more precise for developers.
The latest models, including Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, gain this capability. Unlike the static processing that has been the default since native video analysis launched in 2025, the system now autonomously decides which sections of a video to analyze, at what speed, and using which modality, such as frames, audio, or transcript. This allows it to detect momentary state changes or cuts shorter than one second that might be missed with traditional sampling.
The efficiency gains are most pronounced with lengthy videos, such as tutorials, lectures, or multi-hour recordings. On Google's benchmarks, including LongVideoBench, Gemini 3.7 Flash with agent-based analysis achieved the highest overall quality, demonstrating the best balance of accuracy and cost efficiency. The model can track specific scenes or anomalies across hours of footage without incurring prohibitive token costs.
This development builds upon the "agentic vision" capabilities first introduced with Gemini 3 Flash in January, which allowed models to write and run code to manipulate and check images. The new agent-based video analysis directly ties the model's reasoning to native video tools, enabling it to use an internal tool to grab only the relevant portions of a video file automatically.
The feature is immediately available through the Gemini API for analyzing uploaded videos and YouTube content. Developers can activate the "agentic" processing mode in their API configuration and will be charged at standard Gemini API token rates with no additional fees. Google has published a Developer Guide for implementation.
Looking ahead, Google plans to integrate these improvements into its consumer products. The agent-based analysis is expected to roll out soon to all Gemini app users on compatible devices. In the coming months, it will also power the "Ask YouTube" feature on the playback page, aiming to provide answers more precisely aligned with the visual content of videos.
