Introducing agentic video understanding with Gemini

Google DeepMind is turning on a new way for Gemini to read video, and it’s not just a small tweak. The company says agentic video understanding is now available across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, with access through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

The pitch is simple: stop treating video like a stack of frames at a fixed rate and let the model decide what matters. Instead of grinding through footage at a set 1 FPS, the system can search across frames, audio, and transcripts, then pull in the exact part it needs. Google says that approach cuts token consumption by up to 88%, lowers costs by up to 66%, and improves quality by up to 7% on standard video analysis benchmarks.

That matters most on long videos, where brute-force processing gets expensive fast. Google points to use cases like 10-minute how-to guides, 90-minute lectures, and multi-hour recordings. The company says Gemini 3.7 Flash with agentic understanding gets the best quality and the best quality-to-cost mix among the models it tested, landing on what it calls the accuracy-to-cost pareto frontier.

The useful bit here is not just the savings. Google says the system can help with sub-second moment retrieval, long-form search, anomaly detection, and counting actions or objects over time. In practice, that means things like catching split-second cut points, searching a huge recording without burning through millions of tokens, or resampling a suspicious clip at a higher frame rate.

Google also says developers don’t have to wire all of this together by hand. The model can run an internal tool in an agentic loop to load the right part of a video file, which should cut down development work. Pricing stays on standard Gemini API token rates, with no extra feature fee, and the company says the same efficiency and quality gains will roll out to Gemini app users soon, then later to YouTube’s Ask YouTube feature on the watch page.