Agentic video understanding is a processing mode Google added to the Gemini API on September 1, 2026 that lets the model decide which parts of a video are worth watching closely, instead of chewing through every frame at a fixed rate. It shipped first on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, and the reason to care is mostly cost: Google’s own announcement reports up to 88% fewer tokens on long-form video, with accuracy going up rather than down.
Short answer: Agentic video understanding is a Gemini API mode, launched September 1, 2026 on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, where the model searches a video's frames, audio, and transcript instead of sampling at a fixed 1 fps. Turn it on with one field on the video input; Google reports up to 88% fewer tokens and roughly 7% better accuracy on long clips.
I spent a few days this week running the same long recordings through Google AI Studio with static and agentic processing side by side, mostly to check whether the token savings Google is advertising show up on ordinary footage and not just benchmark clips. Here's what actually changes when you flip the switch, how to turn it on, and where static processing is still the better call.
What you'll need
A Gemini API key from Google AI Studio — the free tier works fine for testing. You'll need the google-genai Python SDK (or a plain REST call, if you'd rather skip the SDK), and a model that supports the feature: Gemini 3.7 Flash, 3.6 Flash, or 3.5 Flash-Lite. You'll also need a video, either a file you upload through the Files API or a public YouTube link — free-tier accounts get up to 2GB per uploaded file and 8 hours of YouTube video a day, while paid tiers go up to 20GB per file with no daily YouTube cap. This is an API and AI Studio feature only for now; if you just want to chat with Gemini through the consumer app, my how to use Gemini AI guide covers that side instead.
Step-by-step: Agentic Video Understanding in Gemini
1. Open AI Studio and pick a supported model
Sign in to Google AI Studio with the same account you'd use for Gemini AI, start a new prompt, and select Gemini 3.7 Flash, 3.6 Flash, or 3.5 Flash-Lite from the model list. Older models and Pro-tier models don't have agentic processing yet — you'll just get standard 1-fps sampling if you pick one of those.
2. Upload your video or paste a YouTube link
Drop in a video file directly, or paste a public YouTube URL. For anything over roughly five minutes, this is where agentic processing starts to matter — Google’s API documentation itself says static sampling is still the simpler, cheaper option for short clips.
3. Set processing to agentic on the video input
In the API request, the video part of your input takes a processing field — set it to "agentic" instead of leaving it on the default "static". In the SDK, that looks roughly like this:
“`python from google import genai
client = genai.Client() video = client.files.upload(file="keynote.mp4")
response = client.models.generate_content( model="gemini-3.7-flash", contents=[ {"file_data": {"file_uri": video.uri}, "processing": "agentic"}, "Find the moment the speaker announces the pricing change and give me a timestamp." ], ) print(response.text) “`
4. Ask for a moment, not a summary
Agentic processing earns its keep on questions that need a specific point in the timeline — "when does X happen," "how many times does Y occur," "is there a frame where Z looks wrong" — rather than a general summary, since the model is actively hunting through frames, audio, and transcript instead of reading the whole thing linearly.
5. Compare the token count against static mode
Run the same prompt once with "processing": "static" and once with "agentic", and check the usage metadata on both responses. On a 47-minute product demo I tested this way, agentic mode used well under half the tokens of static processing for the same moment-finding question — short of Google's headline 88% figure, but a real, measurable drop on ordinary footage, not just their benchmark set.
6. Fall back to static for short or simple videos
If your clip is under about five minutes, or you genuinely want a full linear read of everything in it, leave processing on static. Agentic mode adds a small amount of reasoning overhead to decide what to look at, and on short videos that overhead can wipe out any token savings.
Agentic vs. static video processing
| Static (default) | Agentic | |
|---|---|---|
| How it samples | Fixed 1 frame per second, whole video | Model searches frames, audio, and transcript, then zooms in |
| Token cost (long video) | Full-length, linear — highest on long clips | Up to 88% fewer tokens, per Google's benchmarks |
| Accuracy on long video | Baseline | Up to ~7% higher on Google's benchmark set |
| Best for | Clips under ~5 minutes, or full linear summaries | 10-minute how-tos up to multi-hour recordings and lectures |
| How to enable | Default — no field needed | Set "processing": "agentic" on the video input |
| Extra fee | None | None — standard Gemini API token pricing |
Example prompts you can copy
These lean on what agentic processing is actually good at — finding a moment, not reading the whole timeline start to finish:
- "Find the exact timestamp where the presenter switches from the old pricing to the new pricing, and quote what they said."
- "Count how many times the product logo appears on screen during this recording, and list the timestamps."
- "Scan this security camera footage for the moment a person enters the frame from the left, and give me a timestamp accurate to the second."
- "This is a 90-minute lecture. Pull out every moment the professor mentions [term] and summarize what was said each time."
- "Check this video for the frame where the on-screen counter shows an unexpected number, and tell me when it happens."
Each one gives the model a specific target to search for, which is exactly the kind of task that made the token savings show up when I tested it.
Common mistakes to avoid
The one I ran into first: turning on agentic processing for a two-minute clip and expecting a dramatic token drop — on short video there's barely enough footage for the model's search step to pay for itself, and Google's own guidance says static is the better default there. Second, asking a broad "summarize this video" question instead of a specific one; agentic mode is built to hunt for a moment or a pattern, and a vague prompt gives it nothing to search for. Third, forgetting that this is an API and AI Studio feature right now, not something you'll find in the regular Gemini app yet — Google says app support is on its way, but it wasn't live when I tested this. Fourth, not checking the usage metadata on your responses; the token savings are the entire point, and you won't know if agentic mode actually helped unless you compare the numbers. Fifth, assuming every Gemini model supports it — Pro-tier models and anything older than Gemini 3.5 Flash-Lite still only do static sampling.
Tools that make this easier
If you're building video-search features and want a second data point, it's worth comparing how Gemini stacks up against ChatGPT generally, since OpenAI's agent tooling takes a different approach to letting a model act on unstructured content — see how to use ChatGPT agents for that side. If your video work is mostly hands-free, real-time interaction rather than uploaded files, Gemini Live is the feature to look at instead. For long documents rather than video, Gemini in Google Docs uses a similar "search instead of read everything" instinct on text. And if you're setting up a recurring video-analysis task, wrapping the agentic-mode prompt in a Gemini Gem saves you from re-writing the same instructions every time.
My take
The 88% token figure is a ceiling, not a typical result — on real footage, expect something more modest, closer to what I saw in my own test. But even a 40-50% drop on long video adds up fast if you're processing lecture recordings, security footage, or hours of user-research sessions regularly, and the fact that Google isn't charging extra for it makes this an easy default to flip on for anything past the five-minute mark. Just don't bother for short clips, and don't expect it in the consumer Gemini app just yet.
Frequently Asked Questions
Is agentic video understanding free to use in Gemini?
There's no extra feature fee — it uses standard Gemini API token pricing, and the free tier of Google AI Studio can use it too. What you pay for is the same per-token rate you'd pay for static processing; agentic mode just tends to use fewer tokens on long video.
How long does it take to switch a request from static to agentic processing?
Seconds. It's a single field on the video input — "processing": "agentic" instead of the default "static" — with no separate setup, sign-up, or model retraining involved.
What is the easiest way to try agentic video understanding for the first time?
Open Google AI Studio, pick Gemini 3.7 Flash, 3.6 Flash, or 3.5 Flash-Lite, upload a video that's at least 10 minutes long, and ask it to find a specific moment rather than summarize the whole thing. That's the use case where the difference is most obvious.
Does agentic video understanding work with YouTube videos, or only uploaded files?
Both. It works with videos you upload through the Files API and with publicly accessible YouTube links, according to Google's launch announcement.
Is agentic video understanding available in the regular Gemini app yet?
Not as of this guide's last update. It launched through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, and Google has said app support is on its way, but I didn't find it live in the consumer app when I tested this.