Google this week released agentic video understanding for three Gemini Flash models, a processing mode that lets the model decide which parts of a video file to inspect rather than ingesting every frame at a fixed rate. On the benchmark Google published, one long-video test fell from 397,600 tokens per query to 47,700, an 88.0% reduction, while accuracy on that same test moved from 87.5% to 88.5%. The company put the cost saving at up to 66% and the accuracy gain at up to 7%, figures drawn from different benchmarks in the same set.

The post, dated September 1, 2026, was written by Rohan Doshi, Senior Product Manager at Google DeepMind, and Mario Lucic, Research Director at Google DeepMind. It covers Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, and the feature is live from the same date for uploaded video and YouTube URLs through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

What Google measured

The published bar charts cover three benchmarks: Minerva, described as complex reasoning, and two long-video understanding tests, 1H-VideoQA and LVBench. Each is run twice, once with Gemini 3.7 Flash in standard processing and once with the agentic mode switched on.

On tokens per query, the standard configuration consumed 80,900 on Minerva, 397,600 on 1H-VideoQA and 300,300 on LVBench. With agentic processing those figures fell to 33,600, 47,700 and 36,000 respectively, savings of 58.4%, 88.0% and 88.0%. The headline claim of up to 88% therefore rests on the two long-video tests. On the complex reasoning benchmark, the saving is roughly a third smaller.

Accuracy moved in the other direction, though unevenly. Minerva rose from 73.7% to 79.0%, which Google labels a 7.2% relative gain and which amounts to 5.3 percentage points. LVBench rose from 85.1% to 88.6%, a 4.1% relative gain. On 1H-VideoQA the improvement was 87.5% to 88.5%, a 1.1% relative gain. The single benchmark producing the largest efficiency saving is the one producing the smallest accuracy improvement, and the benchmark producing the advertised 7% accuracy gain is the one where token savings are weakest. Both headline numbers are real. Neither describes the same test.

Methodology for the token and accuracy charts is given in a single line: high thinking level, low media resolution, and 1 FPS for static processing. That last parameter matters, because the comparison baseline is a model sampling one frame per second. A developer already running a lower sampling rate, or already pre-filtering footage, would be comparing against a different starting point.

A second chart plots accuracy against cost per query on 1H-VideoQA and positions Gemini 3.7 Flash with agentic processing at roughly 90% accuracy for close to $0.10 per query. The same model in standard processing sits near 87% at approximately $0.25. Google places four competitor configurations on the same axes: GPT 5.6 Sol at about 80% accuracy and $1.40 per query, GPT 5.6 Terra at about 80% and $0.60, Claude Opus 5.0 at about 65% and $0.50, and Grok 4.6 at about 60% and $0.15. The methodology note records that GPT 5.6 Sol and Terra were run at extra-high reasoning and low detail, while Grok 4.6 and Claude Opus 5 were run at 768 by 768 image resolution with default thinking. Those are configuration choices made by the party publishing the chart, on a benchmark the same party selected.

The gap between tokens and cost

Google states two different reductions in the same sentence: analysis costs down by up to 66% and token consumption down by up to 88%. No chart in the post shows a 66% figure, and the post does not explain the arithmetic connecting the two.

The gap is not trivial. If input tokens fell 88% and cost fell 66%, something in the remaining bill is not scaling with input volume. Output tokens are one candidate, since reasoning traces are billed at a higher rate than input on the Flash line: Gemini 3.6 Flash arrived on July 21, 2026 at $1.50 per million input tokens and $7.50 per million output tokens, with 3.5 Flash-Lite at $0.30 and $2.50. An agentic loop that searches, scans and re-inspects segments generates more reasoning steps than a single pass over a fixed frame sequence, and each step carries its own output. A model that reads less and thinks more will show a token reduction larger than its cost reduction. Google does not say this. It also does not rule it out, and the absence of a per-benchmark cost table leaves buyers unable to reconstruct the figure.

How the loop works

Standard processing ingests video at a fixed frames-per-second rate, defaulting to one frame per second and adjustable through the API. Every frame in that sample enters the context whether it carries relevant information or not. For a ninety-minute lecture, the count reaches five and a half thousand frames before the model has answered anything.

Agentic processing replaces the fixed sample with a decision. Gemini determines what to watch, at what speed, and through which modality, choosing among visual frames, audio and transcript, then fetches only the segments and signals the query requires. The mechanism is an internal tool the model invokes inside a loop to load the relevant portion of the video file. Google positions this as parallel to agentic vision, which pairs code execution with native image understanding.

The company is explicit that this is not a new capability so much as a packaged one. Developers could previously build the same navigation by hand, slicing footage, sampling selectively and stitching the results. What changes is that the orchestration now sits inside a single API call, which Google frames as a reduction in development overhead rather than a capability that did not exist.

Four use cases are named. Sub-second moment retrieval targets split-second state changes and tight cut boundaries that a one-frame-per-second sample misses, which Google links to automated video editing. Long-form needle-in-a-haystack search answers queries across multi-hour footage without consuming millions of tokens. Anomaly detection resamples selected time windows at a higher frame rate to inspect rapid motion and subtle visual artefacts. Counting action and object tracks repeated physical movements and distinct objects over time.

Availability, pricing and the surfaces still to come

Activation requires setting the processing parameter to "agentic" in the API configuration. The published Python sample passes a YouTube URI alongside a text prompt to the model string gemini-3.7-flash through the client interactions endpoint, the same single-request structure Google extended with background execution and remote Model Context Protocol support on July 7, 2026.

Pricing follows standard Gemini API token rates with no additional feature fee. That is a meaningful detail for anyone modelling unit economics, because it means the saving is entirely a volume effect rather than a discounted rate.

Two consumer rollouts are flagged without dates. The feature will reach all users in the Gemini app across Flash and Flash-Lite models, described only as soon. In the coming months it will also power Ask YouTube on the video watch page, which Google says will produce higher-quality answers grounded in the visuals. Ask YouTube drew engagement from more than 140 million users on the watch page during June 2026, according to figures Sundar Pichai gave on the second-quarter earnings call. Gemini 3.7 Flash itself entered the AI Mode model picker for Google AI Pro and Ultra subscribers on August 14, 2026, English only.

One partner on the record

Google states that many early access partners saw strong performance during testing, then publishes a single testimonial. Ibrahim Syed, Founding Engineer at Ponder, is quoted: "At Ponder, we've built our own agentic video-understanding pipeline on Gemini to find usable moments in raw footage. Google's Agentic Video Understanding brought that navigation into a single call, matching our recall while using roughly 3.5x fewer input tokens."

The figure Syed gives, roughly 3.5 times fewer input tokens, works out to a saving near 71%. That is below the 88% ceiling in Google's own charts and above the 58.4% floor. It is also the only third-party number in the post, and it comes from a partner with early access rather than an independent evaluation. No other partner is named, and no aggregate is given for the group Google describes as many.

Why video analysis cost matters to advertising

The advertising relevance is not the feature itself. It is the price of scanning video frame by frame, and who has been paying it.

Contextual video products already run on frame-level analysis. Magnite integrated Anoki's ContextIQ into SpringServe in June 2025 to offer scene-level contextual targeting in connected television, processing visual, auditory and contextual elements together. LG Ad Solutions extended its KERV.ai partnership across four new regions on August 25, 2026, built on frame-level object recognition that feeds both contextual targeting and shoppable overlays. In both cases the economics depend on how expensive it is to look at every frame.

Regulation has pushed in the same direction. The Media Rating Council issued a policy in October 2025 restricting verification vendors from calling their tools brand safety measurement unless those tools examine actual image, video and audio content rather than relying on domain-level or keyword-level analysis. That policy raised the compute floor for an entire vendor category. A processing mode that cuts input volume on long-form video by more than half changes the cost of compliance for anyone building against it, whichever model provider they use.

The token bill has become an operating line rather than a research expense. Agents on the OpenRouter network consumed close to five times as many tokens as human users in data published on August 21, 2026, with consumption up roughly fourteenfold since February. Omnicom reported third-party service costs of nearly $2.9 billion in the first half of 2026 against $1.7 billion a year earlier, and holding companies have started reselling model capacity to clients at a markup. KPMG research published in August 2026 found that 49% of organisations cut agent rollouts when costs outran value. Against that backdrop, a claimed 88% reduction in input volume on the most expensive modality is a commercial argument aimed squarely at the buyers who abandoned deployments on cost.

Whether it converts is a separate question. TripleLift research found that 67% of advertisers cited lack of trust in output as their leading concern about AI in creative work, ranking ahead of brand safety and technical error. Cheaper video analysis does not answer a trust objection. It answers a budget one, and those are not the same constraint. The broader shift toward agentic AI in advertising has consistently produced more announcements than transacted volume.

What the post does not disclose

Several gaps are worth marking. No absolute cost figures accompany the pareto chart beyond an axis running from $1.40 to zero, so the $0.10 and $0.25 per-query readings are estimates taken from plotted points rather than published values. No date attaches to either consumer rollout. No latency comparison appears anywhere in the post, which matters because an agentic loop making multiple passes over a file is doing sequential work that a single fixed-rate ingest is not. Benchmarks were selected and run by Google, competitor configurations were chosen by Google, and no independent replication is cited. The acknowledgement section credits Sergi Caelles, Filip Pavetic, Ahmet Iscen, Suhas Yogin and the Agentic Vision team.

Timeline

Summary

Who: Google DeepMind, through Rohan Doshi, Senior Product Manager, and Mario Lucic, Research Director. Ibrahim Syed, Founding Engineer at Ponder, is the only external voice quoted. Sergi Caelles, Filip Pavetic, Ahmet Iscen and Suhas Yogin are credited alongside the Agentic Vision team.

What: A processing mode called agentic video understanding, activated by setting the processing parameter to "agentic" in the Gemini API configuration. Instead of ingesting video at a fixed frame rate, the model invokes an internal tool to search, scan and inspect selected segments across frames, audio and transcript. Google reports token reductions of 58.4% on Minerva, 88.0% on 1H-VideoQA and 88.0% on LVBench against Gemini 3.7 Flash in standard processing, with accuracy moving from 73.7% to 79.0%, 87.5% to 88.5% and 85.1% to 88.6% on the same three benchmarks. The company separately claims cost reductions of up to 66% without publishing the supporting figures.

When: September 1, 2026. The API capability is live from that date. The Gemini app rollout is described as soon, and the Ask YouTube integration as coming months, with no dates attached to either.

Where: The Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, covering uploaded video files and YouTube URLs, across Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Consumer distribution is planned for the Gemini app and the YouTube watch page.

Why: Long-form video is the most expensive modality to analyse at scale, and the cost has become a live constraint on deployment across advertising technology, where frame-level analysis now underpins contextual targeting, shoppable creative and brand safety verification. Cutting input volume on that modality without a rate change is an argument aimed at buyers who abandoned agent projects on cost rather than capability, though the post publishes no cost table, no latency comparison and no independent verification of its benchmarks.