Google adds agentic video understanding to Gemini Flash models

Google has launched agentic video understanding across three Gemini models: 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. The feature is available now for video uploads and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Instead of the current 'static' method, where a model ingests video at a fixed frame rate (1 frame per second by default, adjustable via the API), agentic video understanding lets the model take an active, goal-directed role: it decides what to watch, at what speed, and through which modality (frames, audio or transcript), and fetches only the segments it needs by invoking an internal tool that loads the relevant part of the video file. Google says this works like its existing agentic vision approach, which pairs code execution with native image understanding, but applied to video. On standard video analysis benchmarks, agentic video understanding cuts analysis costs by up to 66% and token consumption by up to 88% compared with static processing, while improving accuracy by up to 7%, across the three supported models. Google highlights Gemini 3.7 Flash specifically: with agentic understanding it drops token consumption by up to 88% and boosts accuracy by up to 7%, and the company says this puts 3.7 Flash at the accuracy-to-cost pareto frontier among the models it tested for video understanding, meaning no other tested model beats it on both dimensions at once. The efficiency gains are most pronounced on long-form video, from 10-minute how-to guides to 90-minute lectures and multi-hour recordings, where static processing forces developers to choose between high token costs and techniques that drop detail. Google lists four use cases the feature unlocks: sub-second moment retrieval for precise automated video editing, needle-in-a-haystack search across multi-hour videos without burning millions of tokens, anomaly detection by resampling interesting time windows at a higher frame rate, and accurate counting of repeated movements or distinct objects over time. The company says early access partners tested the feature with strong results, though it does not name the partners or quote them directly in the published text. Pricing uses standard Gemini API token rates with no extra fee for the feature; developers turn it on by setting processing to 'agentic' in the API configuration, illustrated with a short Python code example calling client.interactions.create on the gemini-3.7-flash model. Google also says it plans to bring agentic video understanding to consumer products: the capability will roll out to all users in the Gemini app across Flash and Flash-Lite models soon, and in the coming months it will power YouTube's 'Ask YouTube' feature on the video watch page. The work is credited to Sergi Caelles, Filip Pavetić, Ahmet Iscen, Suhas Yogin and the Agentic Vision team.

Key facts

Agentic video understanding launches today across Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform
Versus static processing (fixed 1 FPS by default), the feature cuts analysis costs by up to 66% and token consumption by up to 88%, while improving accuracy by up to 7%, on standard benchmarks
Gemini 3.7 Flash sees token consumption drop by up to 88% and accuracy rise by up to 7%, putting it at the accuracy-to-cost pareto frontier among tested models
The model dynamically decides what to watch and through which modality (frames, audio, transcript) rather than ingesting the whole video at a fixed frame rate, aimed at sub-second moment retrieval, long-form search, anomaly detection and object counting
Standard Gemini API token pricing applies with no extra fee; Google plans to bring the capability to the Gemini app and to YouTube's 'Ask YouTube' feature in the coming months

Why it matters

Video is the most token-expensive input type Gemini handles, and static frame-rate processing has forced a blunt tradeoff: sample densely and pay a large token bill, or sample sparsely and miss details. Agentic video understanding replaces that fixed schedule with a model-driven search over frames, audio and transcript, so the system spends compute only where the content demands it. The result Google reports, up to 88% fewer tokens with accuracy up 7% rather than down, is a case where an efficiency measure also improves quality rather than trading against it.

Who it affects

Developers building on the Gemini API in Google AI Studio or the Gemini Enterprise Agent Platform get the feature immediately for uploaded video and YouTube links. Google says it will later extend the underlying improvements to all users of the Gemini app on Flash and Flash-Lite models, and to YouTube's 'Ask YouTube' feature on the video watch page, which would put agentic video understanding in front of a consumer audience Google describes as billions of users across its products.

How to use it

The feature works with Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite through the Gemini API. Developers enable it by setting the processing field to "agentic" in the API request, as shown in Google's example calling client.interactions.create with a video input and a text prompt. It bills at standard Gemini API token rates; Google states there is no additional fee for using the agentic mode itself. A developer guide is linked from the announcement for further setup details.

How solid is it

The percentage gains, up to 66% lower cost, up to 88% fewer tokens, up to 7% higher accuracy, come from Google's own testing on what it calls 'standard video analysis benchmarks,' without naming which benchmarks or datasets were used, and without giving the absolute baseline cost or token counts the percentages are measured against. The pareto-frontier claim for Gemini 3.7 Flash is likewise Google's own comparison among the models it tested, not an independent evaluation. Google says early access partners saw strong results in testing, but the announcement does not identify those partners or include their testimonials in the published text.

Risks and caveats

All performance figures are self-reported by Google, using unspecified internal benchmarks and undisclosed baselines, so the actual numbers a given application sees will depend on the video content and task. "Up to" figures describe a best case rather than a typical one. No launch date is given for the promised rollout to the Gemini app or to YouTube's 'Ask YouTube' feature beyond "soon" and "in the coming months," so consumer-facing availability timing is unconfirmed.