Video is among the most expensive inputs you can hand an AI model. On September 1, 2026, Google announced a way to cut that bill: agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Instead of swallowing a video frame by frame at a fixed rate, the model now decides for itself what to watch, at what speed, and whether to look at frames, listen to audio, or read the transcript. Google says the result is higher accuracy with up to 88 percent fewer tokens and up to 66 percent lower cost for video analysis.
The announcement landed on the same day Anthropic shipped Claude Fable 5.1 and drew a fraction of the attention, but for anyone who works with long recordings, lectures, meetings, or footage, it may be the more practical release of the week. Here is how it works, what the numbers really say, how to switch it on, and what it means once it reaches the Gemini app.
What Google Announced
Agentic video understanding is a new processing mode for the Gemini models in the Flash family. It is available today through the Gemini API for uploaded videos and YouTube links, enabled by a single configuration setting, at standard token pricing with no extra fee. Google says it is rolling out to all Gemini app users on Flash and Flash-Lite models soon. Early-access partners named in the launch include Ponder, Revyl, Mosaic, and Resemble.AI.
Google’s own summary of the gains: up to 88 percent fewer tokens, up to 66 percent lower cost, and up to 7 percent higher accuracy on long video, with the flagship Gemini 3.7 Flash described as sitting at the accuracy-to-cost frontier among the models Google tested.
Static vs Agentic: How It Actually Works
The old way: one frame per second
Until now, sending a video to a language model meant static processing: the model ingested the file at a fixed frame rate, one frame per second by default and adjustable through the API, turning every second of footage into tokens whether it mattered or not. A 90-minute lecture became hundreds of thousands of tokens, most of them a professor standing still. Developers were forced into a bad trade: pay for everything, or sample fewer frames and risk missing the one moment that answers the question.
The new way: an investigator, not a stenographer
In agentic mode, the model takes a goal-directed role. Given a question, it scans the video, decides which segments are relevant, chooses how fast to move through the rest, and picks the cheapest modality that can answer: the transcript for what was said, the audio for how it was said, the frames only when it needs to see something. It fetches just the segments it needs. As the tech site OfficeChai put it, the model behaves more like an investigator than a stenographer.
The Numbers, and the Fine Print
Where the 88 percent comes from
The biggest savings appear on long-form video, from 10-minute how-to guides through 90-minute lectures to multi-hour recordings. On the long-video benchmarks 1H-VideoQA and LVBench, Google reports token reductions of 88 percent while accuracy lands in the high 80s, with typical token counts on long videos falling from 300,000 to 400,000 down to under 50,000. That is the number the announcement leads with, and on that kind of content it appears to be real.
What “up to” is hiding
The community caught the qualifier immediately. One widely liked reply claimed that a worked example in Google’s materials shows token usage falling from 127,000 to 95,000, a 26 percent cut rather than 88; that figure could not be independently verified, but the underlying point stands: savings scale with video length and with how sparse the relevant moments are. On the complex-reasoning benchmark Minerva, Google reports a 58 percent token saving alongside about a 7 percent relative improvement in accuracy. These are three benchmark results, not a typical figure for everyday clips, and Google’s own qualifier should be taken literally.
Accuracy versus cost against rivals
Google’s launch chart plots accuracy against cost per query on 1H-VideoQA for Gemini 3.7 Flash, OpenAI’s GPT-5.6 Sol and Terra, Anthropic’s Claude Opus 5, and xAI’s Grok 4.6, with Gemini shown at the highest accuracy and lowest cost of the group. That is a striking picture, and it is also Google’s own methodology on Google’s own chart, so it should be read as a vendor benchmark until independent testers reproduce it.
What You Can Do With It
| Use case | What agentic mode changes |
|---|---|
| Finding a moment in a long video | Locating the exact segment with sub-second precision, useful for editing and highlights |
| Needle-in-a-haystack search | Locating one event across multi-hour recordings without processing every frame |
| Anomaly detection | Dynamic frame-rate resampling: slow down where something changes, skip where nothing does |
| Counting repeated actions or objects | Higher accuracy on tallies that static sampling tends to miss |
| Lecture and meeting review | Answering questions from the transcript and audio, pulling frames only when needed |
For non-developers the most relevant row is the last one. Summarizing a recorded meeting or a two-hour lecture, then asking follow-up questions about specific moments, is exactly the task that used to be prohibitively expensive at scale.
A worked example
Picture a two-hour recorded lecture and the question: where did the professor derive the final formula? Static processing turns all 7,200 seconds into frames, tokenizes every one, and then reasons over the whole pile. Agentic processing reads the transcript first, finds the stretch where the derivation is discussed, speeds through the surrounding minutes, and pulls full frames only for the whiteboard moments it needs to see. The answer arrives with a timestamp, and the model has looked closely at a few dozen seconds instead of two hours. Multiply that by a semester of lectures, or a year of team meetings, and the economics change completely.
How to Turn It On
There is no new product to buy. In Google AI Studio or on the Gemini Enterprise Agent Platform, developers set the video processing option in the API configuration to agentic and send the video as usual, either an uploaded file or a YouTube URL. Billing stays at the standard per-token rates for each model, which for Gemini 3.7 Flash currently means promotional pricing of $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, according to Google’s published rates. Because agentic mode consumes fewer tokens, the cost reduction arrives automatically rather than through a new price list.
Coming to the Gemini App
The API launch is the developer half. The half that matters for most people is the rollout to Gemini app users on Flash and Flash-Lite, which Google says is coming soon without a firm date, and a second consumer route with far bigger reach: Google says that in the coming months the same technology will power the Ask YouTube feature on YouTube’s video watch page, so viewers will be able to question a video directly where they watch it. Google has not described the consumer interface yet, but the API already accepts YouTube links and uploads, so the likely experience is asking Gemini about a recording and getting more accurate answers, with the model skipping to the relevant parts instead of reading the whole thing. AI video summarizers are already a widely searched category; agentic processing is what makes them cheap enough to run on everything you record.
When it arrives, the way you ask will matter more than before. Precise questions let the model skip: ask where a decision was made or when a chart changed, rather than requesting a generic summary. Give it timestamps if you have them, and ask for moments rather than overviews.
Part of a Busy Agentic Week at Google
The video launch was one of several Google releases in the same 24 hours, all pointing the same direction. Google Workspace introduced Google Pics, an image tool for Workspace customers and AI Pro and Ultra subscribers that lets users edit individual objects and refine or translate text inside images. The Antigravity coding environment added a boost command that spends extra tokens to reason more deeply on complex tasks. And the Gemma team reported that its open model now runs twice as fast on Apple Silicon thanks to community optimization. Individually these are incremental; together they show Google shipping agentic behavior into every layer of its stack rather than waiting for a single flagship model.
Why This Matters Beyond Google
Three reasons this release deserves more attention than it received.
First, cost was the barrier. Text is cheap to process and video is not, which is why AI assistants have been fluent with documents and clumsy with recordings. Cutting the token bill by half or more on the content people actually accumulate, meetings, lectures, camera footage, is what moves video from demo to daily tool.
Second, the pattern is spreading. The same week saw Anthropic pitch Fable 5.1 on long-running agentic work and OpenClaw ship agents that operate your apps. Letting the model decide where to spend its attention, rather than processing input mechanically, is becoming the default design across the industry, and video is the modality where that idea pays off most.
Third, Google is competing on efficiency rather than raw scale. Rather than announcing a bigger model, it made its cheapest models better at an expensive task, and it aimed the launch at Flash and Flash-Lite, the tiers most developers and consumers actually touch.
Limits and Open Questions
The benchmark numbers are Google’s, the comparison chart is Google’s, and the partner testimonials were provided by early-access customers, so independent verification is still ahead. Google has not published latency figures for agentic mode, and a model that scans and re-samples may take longer per query than a static pass, a trade-off that matters for real-time use. The app rollout has no date. And savings depend heavily on content: a tightly edited two-minute clip has little to skip, so the advertised percentages will not show up on short videos.
The Skill Underneath the Feature
Agentic video only pays off if you ask it a good question. A vague request to summarize a recording still produces a vague summary; a precise question about what was decided, when a number changed, or where a defect appears is what lets the model skip most of the footage and land on the answer. That is a learnable skill, and it transfers to every AI tool that now reads documents, meetings, and video on your behalf. Coursiv builds it with step-by-step guides, short daily lessons, and hands-on practice with AI tools, designed for busy people without a technical background. Check the official site for current course details and pricing.
What to Watch Next
Watch for the Gemini app rollout date and the Ask YouTube integration, which turn this from a developer feature into a consumer one at YouTube’s scale. Watch for independent benchmarks on long-video tasks, especially against GPT-5.6 Sol, which Google’s chart singles out. And watch whether agentic processing spreads to the Pro tier and to audio-only content such as podcasts and call recordings, where the same attention-directing logic should apply.
The Bottom Line
Agentic video understanding is an efficiency release, and efficiency is what unlocks new uses. The 88 percent figure applies to long recordings on Google’s benchmarks, savings on ordinary content will be smaller and depend on how sparse the relevant moments are, and the accuracy gains are Google’s numbers until someone else confirms them. But the direction is right: video that used to be too expensive to ask questions about could become as searchable as a document.