Google is giving Gemini a more selective way to watch videos, allowing its AI models to decide which parts of a recording actually deserve closer attention instead of processing everything at a largely fixed rate.
The new capability, called agentic video understanding, is launching across Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. It is available for uploaded videos and YouTube content through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
The important change is how Gemini approaches video analysis. Conventional processing can sample a video at a fixed frame rate — Google cites 1 frame per second as its default — which creates an awkward trade-off. Sampling more footage can consume huge numbers of tokens, particularly with hour-long recordings, while sampling less aggressively risks missing brief but important visual details.
Agentic video understanding instead lets Gemini dynamically investigate the footage according to the question being asked. The model can search different portions of a video and choose whether it needs frames, audio or transcript information. It can also revisit relevant sections at different sampling rates when additional visual detail is required.
That approach could make a significant difference for developers building AI tools around large video libraries. Google says its benchmarks show token consumption falling by as much as 88% and analysis costs dropping by up to 66%, while accuracy can improve by as much as 7%. Those are Google’s own benchmark results, so real-world performance will inevitably depend on the video, query and application involved.
Gemini 3.7 Flash reportedly delivers the strongest balance of accuracy and cost among the models Google tested with the feature enabled. There is no separate feature fee, with agentic processing using standard Gemini API token pricing. Developers activate it by setting video processing to “agentic” in the API configuration.
The more interesting applications go beyond simply asking an AI to summarize a YouTube video. Google says Gemini can locate sub-second moments, search for specific events across multi-hour recordings, investigate anomalies by examining suspicious sections at higher frame rates, and count repeated actions or objects more accurately.
That could prove useful for everything from automated video editing and lecture analysis to quality inspection and searching massive media archives. The underlying idea is straightforward: an AI should not necessarily need to examine every second of a two-hour recording with equal intensity to answer a question about something that happened for five seconds.
Google also plans to bring the technology beyond developer tools. Agentic video understanding is expected to roll out across Flash and Flash-Lite models in the Gemini app, while YouTube’s Ask YouTube feature is scheduled to adopt it in the coming months.
For AI video analysis, efficiency is increasingly as important as raw capability. Models can already consume enormous amounts of multimodal information; the harder problem is determining what information is actually worth consuming. Google’s agentic approach attempts to solve that by making Gemini less like a machine watching every frame and more like a viewer who knows when to skip ahead — and when to rewind.
