On September 1, 2026, Google DeepMind introduced agentic video understanding in the Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. Instead of fixed frame sampling at one frame per second, the model now decides on its own which parts of the video, audio, or transcript to load and at what resolution. The company's official blog reports a reduction in token consumption by up to 88%, a decrease in cost by up to 66%, and an increase in accuracy by up to 7% on long videos.
The mechanism works as follows: the model receives a link to the video and the user's query, then plans a sequence of actions—extract the transcript, increase the frame rate in a suspicious interval, or switch to audio. Each step is accompanied by internal reasoning, and the final context is formed only from relevant fragments. As a result, an hour-long lecture that in static mode required more than a million tokens is now processed using approximately 100 thousand.
Evaluation was conducted on the 1H-VideoQA and LVBench benchmarks. In agentic mode, tokens were reduced by 88%, and answer accuracy increased by several percentage points. However, Google does not publish detailed ablation studies showing which specific component—modality selection, adaptive frame rate, or internal planning—contributes most to the quality gain.
Previously, most multimodal models, including earlier versions of Gemini and solutions from other labs, relied on static processing: fixed frame rate and full context loading. This approach guaranteed reproducibility but quickly became inefficient on long videos. Google's agentic method resembles approaches already used in text agents with tools, but transfers them to video while preserving multimodality.
Comparison with parallel work shows that Google is betting on efficiency specifically for long content, while some other labs continue to refine static methods with higher frame resolution. The difference in philosophy is clear: one side minimizes the amount of context, the other maximizes its density.
For developers, this means the ability to build video agents that analyze multi-hour recordings without prohibitive costs. New scenarios emerge—automatic search for key moments in lectures, precise video editing by description, monitoring anomalies in industrial recordings. However, it remains unclear how stably the model handles ambiguous queries when relevant fragments are distributed throughout the video.
Independent verification is currently limited to initial developer reports. The community will likely test behavior on edge cases: videos with rapid scene changes, low-quality audio, or contradictory transcripts. The next interesting works will focus on the robustness of agentic planning and its impact on generalization.
The main takeaway of the development is that video analysis becomes not exhaustive but targeted, and this changes the economics of applications working with long multimedia content.
