Gemini's new video mode watches long clips like a human would, and costs a lot less
Google DeepMind's 'agentic' video setting lets Gemini skip around inside a video instead of watching every frame, cutting token use by up to 88% and improving accuracy by up to 7%.
Key points
- Google DeepMind launched agentic video understanding on Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite on the day of the announcement.
- The feature cuts token use by up to 88%, lowers analysis costs by up to 66%, and improves accuracy by up to 7% on standard video benchmarks.
- Developers turn it on by setting
processingtoagenticin the Gemini API, with no extra feature fee on top of normal token pricing. - The same tech will soon reach the Gemini app and power YouTube's 'Ask YouTube' feature on watch pages.
- It supports both uploaded video files and YouTube links inside Google AI Studio and the Gemini Enterprise Agent Platform.
If you've ever asked a chatbot to summarise a two-hour YouTube video and watched your bill (or your patience) melt, this one's for you.
Google DeepMind just switched on what it calls agentic video understanding inside its Gemini models. In plain English: instead of forcing the AI to sit through every second of a video at a fixed speed, the model now decides for itself which bits to watch.
Think of it like the difference between watching a whole football match and asking a mate who was there to just show you the goals.
What actually changed?
Gemini used to watch videos at a steady one frame per second, like a very patient intern taking a photo every tick of the clock. That works, but it burns through tokens (the tiny chunks of text and image data the model is billed for) and often misses split-second details between snapshots.
The new agentic mode lets Gemini act more like a curious researcher. It scans the video, spots what looks relevant to your question, then zooms in at a higher frame rate, or checks the audio, or reads from the transcript. An AI agent, here, is software that can carry out several steps on its own without you holding its hand. We've been watching this approach develop across 31 stories tagged agentic AI since 14 July 2026, and this is among the most practically useful applications we've seen.
Google says the payoff on standard video benchmarks is real: up to 88% fewer tokens, up to 66% lower cost, up to 7% better accuracy. Gemini 3.7 Flash gets the biggest quality boost.
What can it actually do?
Four use cases stand out in Google's write-up:
- Find a specific moment down to the sub-second, useful for automated video editing.
- Answer a question buried inside a multi-hour recording without swallowing millions of tokens.
- Spot anomalies by resampling suspicious moments at a higher frame rate.
- Count repeated actions or distinct objects across a long clip.
These gains matter most on long-form video, from ten-minute how-to guides to 90-minute lectures, where the old fixed-rate approach forced a choice between high token costs and dropping critical details.
What does it cost, and can I try it?
Agentic video is live today through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Pricing uses normal Gemini API token rates, and there's no extra fee for turning the feature on. Developers flip it on by setting processing to "agentic" in the API call.
Here is how the models line up:
| Model | Best for | Extra feature fee |
|---|---|---|
| Gemini 3.7 Flash | Highest quality, best cost-to-accuracy balance | None |
| Gemini 3.6 Flash | Solid all-round video analysis | None |
| Gemini 3.5 Flash-Lite | Cheapest option for lighter jobs | None |
For the rest of us, Google says the feature will roll out in the Gemini app on Flash and Flash-Lite models soon, and will power YouTube's 'Ask YouTube' box on the watch page in the coming months.
Is there a privacy catch?
One thing worth knowing: if you paste in a YouTube link, that's public content. If you upload your own videos, they go to Google's servers for processing, and consumer Gemini app data can be used to improve Google's services unless you turn that off in your Gemini activity settings. Enterprise API traffic is handled under Google Cloud's usual data terms, which don't use your inputs for training by default.
Family holiday footage: probably think twice. Work meeting recordings: check your company's cloud policy first. Public YouTube clips: go wild.
The honest verdict here is that the cost and accuracy numbers are genuinely good, but the bigger deal is developer convenience. Gemini now does the fiddly seek-and-inspect work that engineers previously had to script themselves.
