Google has introduced agentic video understanding, a new approach that lets Gemini decide which parts of a video deserve closer inspection instead of processing the entire timeline at a fixed rate.
Google says the approach can cut token consumption by up to 88%, reduce analysis costs by up to 66% and improve accuracy by up to 7% on standard video benchmarks, potentially lowering the cost of applications that analyze long recordings.
Traditional video processing, which Google calls static processing, typically samples one frame per second and sends those frames to the model. Developers can change the frame rate, but the model still has to process the video broadly.
The new system works differently. Gemini can search transcripts, frames or audio to identify a relevant time range, then increase the frame rate or inspect other signals when the question requires more detail. It can repeat the process if it needs more evidence before producing an answer.
That could be particularly useful for long videos. Google points to applications including 90-minute lectures, 10-minute how-to videos and multi-hour recordings where processing every frame can become expensive or cause important details to be lost.
The system can also locate split-second changes, detect anomalies, count objects or repeated actions and identify precise video-editing boundaries.
Why the shift matters
The bigger change here is not simply that Gemini can understand video more accurately. It is that the model is beginning to decide how it should examine the video.
That distinction could make video analysis more practical for developers building search, editing, monitoring and research tools. Instead of forcing an application to choose between scanning everything and potentially missing something, Gemini can spend more computing power where the question requires it.
For developers, that could mean lower bills and less engineering work. Google says the internal tool loop replaces workflows developers previously had to build themselves. There is still a reason to use the older approach. Google’s developer documentation recommends static processing for shorter videos, particularly clips under about five minutes where latency and full frame-level coverage can matter.
More Google coverage
- New Google Search AI Mode is ‘Total Reimagining,’ Says CEO Sundar Pichai
- In Major Ruling, Judge Finds Google ‘Willfully Acquired and Maintained Monopoly Power’ Over Digital Ad Market
- Google’s Big Bet on Nuclear Energy: ‘The Race to Power AI-Driven Data Centers is Accelerating’
- Computer History Museum Releases Original AlexNet Code: Why It Matters
Coming to Gemini and YouTube
Google launched the feature through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It supports Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, and works with uploaded videos as well as publicly accessible YouTube videos.
The more interesting expansion may be YouTube. Google says agentic video understanding will power the Ask YouTube feature on video watch pages in the coming months, allowing Gemini to provide answers grounded more directly in what appears in the video.
For developers, the broader shift is toward multimodal AI systems that decide how much data they need before answering. If Google’s efficiency claims hold up in production workloads, that could make video search, monitoring and analysis less expensive to deploy at scale.
Read more: Learn how Google’s Gemini and other AI models fit together and which tools are designed for different AI workloads.