Google adds agentic video understanding to Gemini

Google announced agentic video understanding for Gemini on September 1, 2026, a new way for the model to analyze video by dynamically scanning segments instead of sampling frames at a fixed rate.

What's new

Instead of processing video at a constant frame rate regardless of content, the new approach lets Gemini adjust how densely it samples a clip based on what's actually happening — spending more attention on fast-changing moments and less on static ones. Google says this enables sub-second moment retrieval for precise video editing, more efficient analysis of long-form content without blowing through the token budget, better anomaly detection through dynamic resampling, and more accurate counting of actions and objects across a clip.

The headline numbers: "Reduce analysis costs by up to 66% and token consumption by up to 88%, while improving accuracy by up to 7%." The capability is live now on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Google says "the feature is available today for video uploads and YouTube videos via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform," with a rollout to Gemini app users underway and integration planned for YouTube's "Ask YouTube" feature.

Context

This lands alongside a run of other Gemini video work this year — Gemini Omni's multimodal video generation and the promotion of Gemini 3.1 Flash Image and 3 Pro Image to general availability with video-to-image generation. Where those releases were about generating video and images, agentic video understanding is about the other direction: getting a model to reason accurately and cheaply about video that already exists, at YouTube's scale of long-form content.

Why it matters

Fixed-frame-rate sampling has been one of the more mundane but real cost problems in video LLMs: a model burns tokens on redundant near-identical frames during a static shot, then still might miss a fast action in between sampled frames. A model that adjusts its own sampling density fixes both ends of that problem at once, which is why the improvement shows up simultaneously in cost, token count, and accuracy rather than trading one for another. Cheaper, more accurate video understanding has direct implications for video editing tools, content moderation pipelines, and any agentic workflow that needs to ground itself in what's actually happening on screen — and folding YouTube's "Ask YouTube" into the same capability suggests Google intends this as infrastructure across its own consumer video surface, not just an API feature.