Why does AI still struggle with a 90-minute video?
If you have ever pasted a long webinar or training recording into an AI tool and got back a summary that missed the one moment you cared about, the problem was not your prompt. Until this month, Gemini watched video the way a security camera records: one frame every second, start to finish, whether or not the question needed it. Agentic video understanding, launched by Google DeepMind on 1 September 2026, changes that. The model now decides which parts of the video to inspect, at what speed, and through which signal.
Agentic video understanding is a Gemini processing mode in which the model treats a video like a searchable file rather than a fixed stream. It reads the transcript first, locates the relevant time range, then loads only those frames or audio segments, re-inspecting at a higher frame rate when a question demands detail.
The old behaviour is called static processing. According to Google's token documentation, static mode costs roughly 100 tokens per second of video at low resolution and about 300 tokens per second at high resolution. A one-hour lecture therefore costs around 1.08 million tokens before you have asked a single question. Google's launch post gives the agentic figure for the same lecture as roughly 108,000 tokens, depending on the prompt.
For a marketer reviewing competitor launch events, an educator checking a recorded lecture, or an ops manager pulling decisions out of a two-hour town hall, this is the difference between a tool you use once and a tool you use every week.
What did Google actually launch on 1 September 2026?
Google DeepMind launched agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Google reports up to 88% fewer tokens, up to 66% lower analysis cost and up to 7% higher accuracy on standard video benchmarks. It works on uploaded files and public YouTube links, at standard token pricing with no feature fee.
The announcement was written by Rohan Doshi and Mario Lučić of Google DeepMind. Three points in it matter for a practitioner rather than a developer.
---The feature is a mode, not a new model. You keep using the Flash models you already know and switch video processing from static to agentic.
---It is live today in Google AI Studio, which is a free browser interface, not just the raw API. Google says a Gemini app rollout for Flash and Flash-Lite models will follow "soon", and that the same engine will later power YouTube's Ask YouTube feature.
---The efficiency claims are maximums from Google's own benchmarks. MLQ News, reporting on 4 September, noted that Google's developer documentation now also lists Gemini 3.8 Flash as supported, so the model list is already wider than the launch post.
The four capabilities Google highlights are sub-second moment retrieval, needle-in-a-haystack search across multi-hour recordings, anomaly detection by re-sampling a short window at higher FPS, and counting repeated actions or objects. Notice that all four are questions about a specific part of a video, not requests to summarise everything.
How does agentic video understanding work in plain terms?
Static processing feeds every second of a video into the model's context at a fixed rate, so a long file is expensive before any reasoning starts. Agentic processing adds a navigation loop: Gemini scans the transcript, picks a time range, requests only those frames or audio, checks whether it has enough evidence, and repeats if needed. Fewer frames enter the context, so cost falls and precision rises.
Think of the difference between reading a 300-page report cover to cover and using the index. Static mode reads every page. Agentic mode looks up the topic, jumps to page 142, reads the surrounding paragraphs closely, and goes back to the index if the answer is incomplete.
Google's documentation describes three token buckets in agentic mode: navigation reasoning, on-demand tool-use tokens for the segments the model fetches, and the final output. This is why the bill depends on your question. A narrow question about one moment fetches very little. A request to describe every segment sends the model back through the whole timeline, and the saving shrinks accordingly.
Developers Digest, in a 1 September analysis, put it precisely: this is a routing claim, not a compression claim. The 88% figure is real when your question points at a fraction of the footage. It is not a discount on watching everything.
How do you use it in Google AI Studio without writing code?
Open Google AI Studio, pick Gemini 3.7 Flash, attach a video file or paste a public YouTube URL, and set video processing to agentic (the API setting is processing: "agentic"). Then ask a targeted question with a timestamp, a named speaker or a specific event. Free-tier accounts can upload files up to 2 GB and process up to eight hours of YouTube video per day.
The step that most people get wrong is the question itself. Agentic mode rewards specificity because specificity tells the model where to look. Compare these two prompts on a 75-minute recorded product webinar.
Weak prompt: "Summarise this webinar."
Strong prompt, ready to copy:
Try this prompt:
"You are reviewing a 75-minute recorded product webinar for a marketing team. Do the following in order. (1) Find every moment where the presenter states a price, a launch date or a named integration partner, and give me the timestamp plus a one-line quote for each. (2) Locate the Q&A segment and list each audience question with the presenter's answer in under 25 words. (3) Identify the single slide that stayed on screen longest and describe what it shows. (4) Flag anything you could not verify from the transcript and had to infer from visuals. Return the result as a table with columns: Timestamp, Type, Content, Confidence."
Every instruction in that prompt gives the navigation loop a target: a type of statement, a segment, a visual property. Instruction (4) is the one practitioners skip and should not, because it forces the model to separate what it heard from what it inferred.
If you have already set up recurring work in ChatGPT using the approach in our scheduled tasks guide, the same principle applies here: define the output shape before you run, and the tool becomes repeatable instead of a one-off experiment.
Which video tasks benefit most, and which should stay in static mode?
Agentic mode wins on long recordings and targeted questions: finding one statement in a two-hour meeting, counting how many times a demo failed, or pulling every timestamp where a competitor is named. Static mode remains the better choice for clips under five minutes, for tasks that need every frame covered, and for outputs where strict formatting matters more than token savings.
Google's own documentation says agentic navigation can increase time to first token on clips shorter than five minutes, and recommends static processing for latency-sensitive short videos and for full-frame coverage tasks.
The most useful outside evidence so far is a small benchmark published on the Google AI developer forum on 3 September by PaperEdits, which disclosed a commercial affiliation. On six synthetic ten-minute videos with frozen prompts and no retries, agentic mode recovered 18 of 20 brief events versus 15 of 20 for static mode, and scored 0.68 against 0.55 on an edit-decision F1 measure. But static mode used 26.42% fewer tokens on that particular test, cost 23.01% less, scored slightly higher on broad moment retrieval, and one of six agentic outputs failed the required JSON format.
Read that as a decision rule rather than a contradiction. Short video plus broad question: stay static. Long video plus narrow question: go agentic. Six synthetic videos do not settle anything, but they match Google's own guidance on where the mode is designed to help.
Practical scenarios where the maths clearly favours agentic mode:
---A content marketer scanning a competitor's 90-minute keynote for every pricing mention.
---A trainer checking a three-hour recorded course for the exact moment a safety step was skipped.
---An ops manager extracting every decision and owner from a two-hour town hall, with timestamps for follow-up.
---A social media manager finding the three tightest cut points in a long interview for short-form clips.
What are the common mistakes and hard limits?
The five failures that waste the most time are asking for a full summary and expecting the 88% saving, testing on a three-minute clip and concluding the mode is slow, exceeding the one-hour default context limit without switching to low resolution, trusting counts and visual claims without a verification step, and assuming the mode is already in the consumer Gemini app when it is currently an AI Studio and API feature.
Some numbers to keep at hand, all from Google's video understanding documentation as reported on 4 September 2026:
---Models with a one-million-token context handle videos up to one hour at default media resolution, or three hours at low resolution.
---File uploads are capped at 20 GB for paid users and 2 GB on the free tier.
---Free-tier YouTube processing is limited to eight hours of video per day.
---Standard pricing for Gemini 3.7 Flash and 3.6 Flash is listed at US$1.35 per million input tokens and US$6.75 per million output tokens through 31 December 2026; Gemini 3.5 Flash-Lite is US$0.54 and US$4.50.
Two softer limits matter as much. First, Google's safety documentation states plainly that video outputs can be inaccurate or biased and recommends human review, so treat counts, quotes and visual descriptions as drafts until checked against the timestamp. Second, this is a multi-step agent loop, and it inherits the reliability profile of agents: more places to fail than a single pass, and occasional structured-output slips like the JSON failure in the PaperEdits test. Ask for a table, then eyeball it.
Model availability is also a moving target. This article reflects the 1 September launch list plus the 3.8 Flash addition visible in documentation on 4 September. Check the current documentation before building a workflow around a specific model name.
Try it now: a 20-minute test on a video you already have
Pick one recording you have been meaning to review, ideally longer than 30 minutes. In Google AI Studio, run the same targeted prompt twice, once in static mode and once in agentic mode, and compare three things: the token count shown in the response, whether each answer includes a checkable timestamp, and how many claims you can verify in five minutes of spot-checking.
A good first test is a recorded webinar you have already watched, because you know what the right answers are. Ask for the timestamps of three moments you remember. If agentic mode finds them with fewer tokens, you have a workflow. If it misses one, tighten the prompt with a speaker name or a visual cue and run again. That single loop teaches you more about the mode than any benchmark.
The bigger lesson is the one underneath every new AI feature this year. The capability arrived as a setting, but the value arrives through the question you ask. A precise brief turns a 90-minute video into a two-minute answer; a vague one turns it into an expensive summary. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Ready to put AI to work on the recordings nobody has time to watch?
Reviewing long videos is exactly the kind of repeatable task an AI employee can own. We'll walk you through every step, from choosing the right model and prompt to setting up a workflow your team can run every week without you.
