Google Gemini's new agent-based video analysis cuts token usage by up to 88 percent

Google is bringing agent-based video analysis to Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite - a change to how the models process video that moves away from the conventional fixed-rate frame sampling approach. Rather than scanning footage at a uniform interval regardless of content, the model now decides for itself which segments are worth examining and at what resolution. Google says the result is a reduction in token usage of up to 88 percent, alongside accuracy improvements that are especially noticeable on multi-hour recordings.

The technical distinction here is meaningful. Fixed-rate frame extraction is simple and predictable, but it treats every second of a video as equally important, which is rarely true in practice. A two-hour recording of a conference, for instance, may have long stretches of near-identical frames punctuated by moments of significant change. An agent-driven approach can allocate attention where it is actually needed, skipping redundant content and zooming in on relevant segments - much like how a human analyst would scrub through footage rather than watching it at 1x speed.

Token usage is a practical concern for anyone building applications on top of large multimodal models. Video has always been expensive to process in this context because high frame counts translate directly into large token counts, which affect both cost and latency. An 88 percent reduction, if it holds across a range of real-world tasks, would substantially lower the barrier to running video analysis at scale - whether for content moderation, sports analytics, surveillance review, or document-heavy workflows involving recorded screencasts.

The feature lands across the Flash tier of Gemini models rather than the more capable Pro line, which suggests Google is positioning this as an efficiency-focused improvement suited to high-volume, cost-sensitive use cases. It also fits into a broader industry pattern of treating inference not as a single forward pass but as a multi-step process where the model exercises some control over its own inputs - an approach that has shown promise in text-based reasoning and is now finding its way into other modalities.