Introducing agentic video understanding with Gemini

Agentic video understanding is launching across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, available today for uploads and YouTube videos through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It uses standard token pricing with no additional feature fee.

The mechanism is the point. Static processing ingests video at a fixed frame rate — one frame per second by default, adjustable through the API. Agentic processing pairs the model's reasoning with native video tools so it can search, scan, and inspect specific segments across frames, audio, and transcripts, fetching only what it needs. Google reports up to 66 percent lower analysis cost, up to 88 percent fewer tokens, and up to 7 percent better accuracy across standard video benchmarks, with the gains concentrated on long-form content from ten-minute how-tos to 90-minute lectures and multi-hour recordings.

Fixed-rate sampling was always a bad compromise. At one frame per second a two-hour recording is 7,200 frames whether the answer lives in ten seconds of it or not, and the workarounds — dropping frames, chunking, pre-filtering — trade detail for budget. Letting the model choose what to watch, at what speed, and through which modality is the obvious fix, and Google notes developers could already do this by hand. What changes is that the loop is internal, so the orchestration overhead disappears.

Signal Desk's read: the efficiency numbers will get quoted and the capability change is what matters. Sub-second moment retrieval, accurate counting of fast movement by rewatching a segment at a higher frame rate, and needle-in-a-haystack search across multi-hour video are not cheaper versions of what static processing did badly. They are jobs that were not economically viable before. Compliance review, sports and security footage, lecture archives, and long-form support content all move from batch experiments to something you can put behind a product.

Pricing is the aggressive part. Standard token pricing, no feature fee, on Flash-tier models. Google is not monetizing the capability; it is using it to make Flash the default place long video gets processed. If the pareto-frontier claim holds up in independent testing, competitors have to answer on cost per analyzed hour, not on model quality.

The caution is that all the published figures are Google's own, measured on standard benchmarks including LongVideoBench, and "up to" is carrying weight in every one of them. Accuracy improving by up to 7 percent is a benchmark statement, not a promise about your footage. Anyone building on this should measure on their own video before assuming the savings.

This follows the same pattern as agentic vision, which combined code execution with native image understanding. The through-line is Google converting passive multimodal ingestion into tool use the model drives itself. Google also says the capability will reach Gemini app users on Flash and Flash-Lite soon, and will power YouTube's Ask YouTube feature on the watch page in the coming months.