Gemini gains agentic video understanding

Google DeepMind introduced agentic video understanding for Gemini on September 1, 2026. Instead of reading a file straight through, the model decides which segment to load, how fast to sample it, and whether to work from frames, audio or the transcript.

From a fixed tick to a targeted search

Static processing takes the video at a fixed frame rate, one frame per second by default and adjustable through the API. On a recording that runs for hours, every second gets the same attention, including the ones where nothing happens.

The agentic mode runs a loop instead. The model calls an internal tool that pulls only the part of the file it needs at that moment, and it can go back over fast movement at a higher frame rate. Google names four uses: locating a moment down to the sub-second level for automated editing, finding a single passage inside hours of footage, spotting anomalies, and counting objects or actions over time.

The numbers come from the vendor

Against static processing, Google reports up to 88 percent fewer tokens, up to 66 percent lower cost and up to 7 percent better quality. For the quality figure the post points to LongVideoBench, a benchmark for long-form video understanding.

All three are ceilings rather than averages. The company says the gains are most pronounced on long-form video, and gives no separate figure for short clips. It does not publish the measurement setup, the set of videos tested, or any spread around the reported values.

Where it runs

The feature is live in the Gemini API through Google AI Studio and on the Gemini Enterprise Agent Platform, covering both uploaded video and YouTube video. Developers enable it by setting the processing parameter to agentic in their configuration; the sample code in the post uses Gemini 3.7 Flash. Billing stays on the standard token price, with no separate fee for the feature.

Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are supported. Google says a rollout to Gemini app users is coming, along with an Ask YouTube feature on video watch pages. Neither has a date in the announcement.

What this story does not establish

Exactly one source was available for this article: Google's own announcement. There is no independent check of the reported savings and no field reports yet. Whether an 88 percent token reduction carries over to other footage cannot be read off the post.

The paper trail is thinner than it looks, too. The link that circulated on deepmind.google redirects to blog.google - the same sender, not a second newsroom. This remains a vendor announcement, and it is treated as one here.