Google quietly rolled out a new way for Gemini to watch video on September 1: agentic video understanding. Instead of decoding a fixed number of frames per second, the model reasons over the video's timeline the way a person scrubbing through footage would. Google says that cuts video tokens by up to 88%, cost by up to 66%, and even nudges accuracy up by up to 7% on its own video benchmarks. It shipped live on Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, no extra fee.
What Is Gemini Agentic Video Understanding?#
Here's the actual mechanic, from Google's own announcement, published September 1. The old "static" mode extracted frames at a flat 1 frame per second and ran a single pass over the whole clip — a 90-second product cutaway got the same treatment as a 90-minute webinar. Agentic mode gives the model native video tools instead: search, scan, inspect. It loops, decides which segments actually matter, and only "watches" those closely.
Turning it on is one line for a developer — "processing": "agentic" on the video part of a Gemini API request. It works on uploaded files and on public YouTube URLs, and as of this writing it's live on Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Standard token pricing applies. No surcharge for the smarter mode, at least not yet.
Does This Change Anything For Your Business?#
At WebEpex we build Meta ad creative and WhatsApp automation for clients across India, the UAE, the UK and the US, and video review has always been one of the slower, pricier things we touch. If you don't run any — no ad creative, no client testimonials, no long-form content sitting unused — this doesn't affect you today. Skip ahead.
If you're running Meta ads with video creative, or sitting on hours of webinar and testimonial footage nobody's repurposed, this is the part worth reading. Watching video with an LLM has always been expensive, because every second decoded eats tokens whether anything happens in that second or not. A ten-minute product demo has been costing roughly what a ten-minute static shot of an empty room costs. Agentic mode is Google admitting that was wasteful, and building a model that skims for what matters instead of decoding everything blindly.
My read: this matters more for ad and content teams than for chatbot builders, since most of our WhatsApp automation work never touches video at all.
What Are the Real Trade-Offs?#
The upside is Google's own number, not mine:
- Up to 88% fewer video tokens per Google's benchmarks
- Up to 66% lower processing cost at standard API pricing
- Up to 7% higher accuracy on video understanding benchmarks
That last point matters — a cost cut that also hurts quality isn't really a cut. But there's a real caveat, flagged in MarkTechPost's coverage of the launch: static mode is still better for clips under five minutes and for frame-by-frame precision work. Looping and reasoning adds latency, so a quick five-second cutaway probably doesn't need this at all. And it's a hosted API feature only — no open weights, no self-hosting — so if a client ever needs everything kept on their own infrastructure, this option isn't on the table for them.
How Is WebEpex Using This?#
That's the first thing we're pointing agentic mode at this week: a batch of public YouTube ad libraries, run in parallel with our manual process. We haven't touched a client deliverable with it yet, so I won't claim a before-and-after number we don't have — same discipline we used building WhatsApp lead response automation, test on our own systems first. If the savings hold up against real ad libraries and not just Google's benchmark clips, competitor-ad research gets cheaper. If not, I'll write that post too.
The same pipeline, if it holds, also feeds the content-repurposing side of AEO — turning a client's existing video into the kind of specific, quotable text an AI answer engine can actually cite.
What Should You Do This Week?#
If you're not spending real money on video ads or sitting on unused footage, do nothing — this genuinely isn't urgent for you. If you are, ask whoever runs your ad account whether they're already using a video model anywhere in reporting or research.
If the answer's no, that's fine too — static-frame tools still work, they're just pricier per hour of footage, the same kind of cost math we walked through on GLM-5.3-Flash's price change this week. The one thing worth doing regardless: pull your best-performing video ad from last quarter and get it transcribed into two or three pieces of written content. That part doesn't need agentic mode. It just needs someone to do it.
If you want to know whether any of this actually changes your ad spend or your content backlog, send me what you're running and I'll tell you straight — no pitch attached. cal.com/webepex/growth-review
