Google DeepMind has introduced a new capability for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite: agentic video understanding. The model no longer pushes the video through a fixed algorithm. It decides on its own what to watch, at what speed and through which modality (frames, audio or transcripts).
This lets the model focus dynamically on the relevant segments of a video, which saves a lot of resources and improves accuracy. Compared to traditional methods, token consumption drops by up to 88% and costs by up to 66%. The quality and accuracy of video analysis go up by as much as 7%. It is available through the Gemini API and Google AI Studio.
Before, analyzing a video meant running the whole stream frame by frame and burning a huge number of tokens even on moments with no useful information. Now Gemini itself "decides" which segments deserve attention. This makes it practical to "find a needle in a haystack" in long videos, for example when searching for specific events or actions.
The first thing I will test with this new capability is how cost and accuracy change for action recognition in video, in the context of security systems or production process monitoring. This directly affects efficiency and ROI for tasks where you previously had to sacrifice either depth of analysis or budget.
Who this is useful for:
- Teams building video content analysis solutions (for example, in logistics, security systems, retail).
- Developers whose agents need to understand what is happening in a video without going through every second.
- Anyone who has hit the cost ceiling of video processing in multimodal scenarios.
Who this matters less for:
- Simple tasks where analyzing static images is enough.
- Teams that already have highly optimized, specialized pipelines and do not need the flexibility of an agentic approach.