Google DeepMind just launched agentic video understanding across its newest Gemini models: Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. According to Google DeepMind, the feature lets Gemini actively search, scan, and inspect video instead of passively watching it frame by frame. The result is faster, cheaper, more accurate video analysis. It’s available today.
What stands out here is the shift in how the model handles video. Until now, Gemini ingested video at a fixed rate, usually one frame per second. Agentic video understanding pairs the model’s reasoning with native video tools, so it can dynamically hunt through visual frames, audio, and transcripts to find exactly what matters.
What it does
Google DeepMind says the new capability unlocks video tasks that static processing struggled with:
- Sub-second moment retrieval: pinpoint a precise moment in a clip rather than scanning the whole thing.
- More accurate anomaly detection: catch the outlier frame or event that a fixed sample rate would skip.
- Precise counting: count objects or occurrences with far fewer errors.
- Targeted inspection: zoom in on specific segments across frames, audio, and transcript together.
The approach mirrors agentic vision, which combines code execution with Gemini’s native image understanding. This is the same idea applied to moving pictures.
The numbers
This is where the launch gets interesting. Across standard video analysis benchmarks, Google DeepMind reports that Gemini models with agentic video understanding:
- Cut analysis costs by up to 66%
- Reduce token consumption by up to 88%
- Improve accuracy by up to 7%
Cheaper and better at the same time is a rare combination. Usually you trade one for the other.
Why long video is the real win
The efficiency gains show up most on long-form content, from 10-minute how-to guides to 90-minute lectures and multi-hour recordings. Google DeepMind points out the old tradeoff developers faced: pay high token costs to process a long video in full, or use shortcuts that drop critical details.
Agentic video understanding sidesteps that choice. The model searches for the relevant segments instead of chewing through every second at a flat rate. For anyone building on top of hours of footage, that changes the math on what’s affordable to run.
How to get it
The feature is live now. Per Google DeepMind, you can use it with both direct video uploads and YouTube videos through:
- The Gemini API in Google AI Studio
- The Gemini Enterprise Agent Platform
The default sampling rate stays at one frame per second and remains adjustable via the API, so developers keep control over how static or agentic they want a given job to be.
Where this fits
Video has been the expensive, awkward corner of multimodal AI. Text is cheap to process, images are manageable, but video multiplies token counts fast, and long recordings made many use cases too costly to bother with. Compressing the cost by two-thirds while nudging accuracy up moves a lot of previously impractical projects into reach.
Think about the practical applications. Security teams scanning hours of camera feeds for a single anomaly. Educators indexing lecture archives so students can jump to the exact 40 seconds they need. Media companies tagging footage libraries. Support teams pulling the right moment out of a screen recording. All of those get cheaper and sharper under this model.
One caveat worth noting: the figures Google DeepMind shares are ceilings, phrased as “up to” 66% cost reduction, “up to” 88% fewer tokens, and “up to” 7% better accuracy. Real results will depend on your video length, content, and task. The gains are clearly biggest on long-form material, so shorter clips may see more modest improvement.
This is significant because it attacks the cost wall that kept video AI on the sidelines for a lot of builders. When the model decides what to look at instead of grinding through every frame, video analysis starts to behave more like search than transcoding. Expect competitors to chase the same agentic approach, and expect a wave of products built on footage that used to be too expensive to touch. Full benchmarks and access details are available at the original source.