Google's Gemini video analysis cuts AI token costs by 88% for businesses

Google's new video analysis tool for Gemini models reduces token usage by up to 88% compared to traditional frame-by-frame scanning, lowering costs by 66% per benchmarks while supporting long-form vid

Google has shipped agent-based video analysis to its Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. The feature dynamically searches video footage instead of scanning frames at a fixed rate, reducing token usage by up to 88% and cutting costs by 66% per Google’s stated benchmarks. It detects moments shorter than one second—including state changes or cuts—and supports videos from 10-minute tutorials to multi-hour lectures. The system is currently available via the Gemini API for finding specific scenes in long videos, with no additional fees. It builds on 'agentic vision' technology previously shipped in Gemini 3 Flash in January 2026.

The mechanism selectively retrieves only relevant video sections, avoiding full-frame processing. This allows the system to handle long recordings efficiently while improving accuracy slightly on benchmarks like 1H-VideoQA and LVBench.

For businesses and consumers, this makes video analysis cheaper and more scalable—potentially expanding access to tools for education, content creation, and real-time video understanding. However, the feature remains API-only for now, with planned rollouts to the Gemini app and YouTube later this year.

What to watch: YouTube integration timelines, real-world accuracy impacts beyond benchmarks, and whether cost savings translate to wider adoption. Google’s claims are based on internal benchmarks, not independent testing, and the feature is not yet available to end users.

Source: The Decoder