Google adds agentic video analysis to Gemini models

Google has launched agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. As detailed in the Google DeepMind Blog, the feature lets Gemini decide which parts of a video need closer inspection instead of processing the entire file at a fixed frame rate.

The capability is available now for uploaded videos and YouTube links through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Developers activate it by setting the API configuration to agentic.

Why it matters

Traditional video analysis typically samples footage at a set rate—1 frame per second by default—creating a trade-off between detail and cost. Agentic processing uses an internal tool loop to search, scan and retrieve relevant sections across video frames, audio and transcripts. Gemini can choose what to watch, how quickly to inspect it and which media signal is most useful for the question.

Google says benchmark testing showed up to 66% lower analysis costs, 88% less token consumption and a 7% accuracy improvement. The savings are particularly relevant for long videos, including lectures, tutorials and multi-hour recordings, where static processing can become expensive or miss brief events.

The approach also expands what builders can automate: finding moments lasting less than a second, searching for a specific detail buried in hours of footage, detecting anomalies by examining selected intervals at higher frame rates, and counting objects or actions more precisely. Google says Gemini 3.7 Flash delivered the strongest overall quality and the best quality-to-cost balance among the tested models.

For teams building video search, moderation, editing or monitoring tools, the main change is less custom orchestration. Rather than writing separate logic to locate and resample relevant scenes, developers can give Gemini a goal and let the model direct the inspection process.

Source: Google DeepMind Blog