Gemini Can Now Watch Videos the Smart Way. Here's Why That Changes Your Build.

Google just shipped agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. According to Google DeepMind, it cuts token consumption by up to 88%, reduces costs by up to 66%, and improves accuracy by up to 7% compared to the previous approach. That combination, lower cost AND better results, is rare enough to pay attention to.

What "Agentic" Actually Means Here

Before this update, Gemini processed video the blunt way: it sampled frames at a fixed rate, typically one frame per second, and churned through every single one whether they were useful or not. Think of it like reading a whole book to answer one question that's buried in chapter nine. You pay for every page.

Agentic video understanding changes the strategy. Instead of sampling everything, Gemini now actively decides what to watch, when to watch it, and whether to focus on the visuals, the audio track, or the transcript. It scans, seeks, and zooms in on the moments that matter. The model is doing the editorial work that a human researcher would do when scrubbing through footage. Less waste, more precision.

This matters most for long-form video. A 60-minute recorded design critique, a 90-minute user research session, a multi-hour conference talk: previously, analysing any of these meant burning through an enormous number of tokens and paying accordingly. Now, the model fetches only what it needs.

The Concrete Change for Designers Who Build

Here is the thesis: video just became a viable input for the apps you are designing and building.

Until now, adding any kind of video analysis to a vibe-coded or AI-assisted app felt like a luxury. The token costs made it impractical for anything longer than a short clip, and the accuracy on longer recordings was spotty. Both of those friction points have shifted.

Think about the kinds of products designers actually want to build. A tool that watches a recorded user-testing session and surfaces the three moments where a participant got confused. An app that scans a client's existing tutorial videos and extracts a structured outline. A portfolio review tool that can pinpoint the exact timestamp where a presenter loses their thread. All of these were technically possible before; none of them were economically sensible at scale. That calculus is changing.

How to Actually Start Using This

The feature is live now in Google AI Studio and the Gemini Enterprise Agent Platform. It uses standard Gemini API token pricing, no extra fee on top.

To turn it on, you pass "processing": "agentic" when you set up a video input in the API call. That single flag is the switch. If you have been experimenting in AI Studio, you can test it directly with a YouTube URL or an uploaded file, no local infrastructure needed.

The four capabilities Google calls out are worth knowing by name so you can match them to your ideas:

• Sub-second moment retrieval: finding a precise instant in a video, even a fast cut or a brief state change. Useful for anything involving automated editing or highlight reels.
• Long-form needle-in-a-haystack search: answering a specific question across a very long recording without processing the whole thing. Useful for research tools or meeting summarisers.
• Anomaly detection: spotting something unusual in footage by resampling the interesting window at a higher frame rate. Useful for quality-check tools or anything monitoring physical processes.
• Counting actions and objects: tracking repeated movements or distinct items over time. Useful for fitness, manufacturing, or spatial design contexts.

Gemini 3.7 Flash is the model Google recommends if you want the best balance of quality and cost. The documentation and developer guide are available on Google AI Studio now.

What to Keep in Mind

A few honest caveats before you go all-in.

The efficiency numbers come from Google's own benchmarks, and real-world results will vary depending on what you are asking the model to do and what kind of video you feed it. The accuracy improvement is described as up to 7%, which is meaningful but not a transformation in kind, more a refinement.

The feature is also rolling out to the Gemini consumer app and will eventually power YouTube's "Ask YouTube" feature on the watch page, which tells you Google is betting seriously on this. But it also means the API behaviour could evolve as the feature matures.

For now, the smartest move is to prototype with a real use case from your own work. Grab a recording you already have, a user interview, a stakeholder presentation, a workshop session, and test what the model actually surfaces. That is more useful than any benchmark.

The door to video-powered tools just got significantly cheaper to walk through. The interesting question is what you will build once you are inside.