Gemini Can Now Scan Entire Videos Itself: Agentic Approach Saves Tokens and Costs

The model chooses where to look

Google has improved video processing for the Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models. It is transitioning to an agentic approach, where the model receives a task and dynamically scans video segments itself until it finds an answer. This means the entire footage chopped into frames doesn't have to be sent to the context window.

An end to wasting context on static scenes

For developers building on video analysis, this changes the economics of the entire process. Google claims this dynamic approach reduces token consumption by up to 88% and real API costs drop by up to 66%. Besides saving money, this positively impacts accuracy, as the model doesn't get overwhelmed with unnecessary noise from static passages and focuses only on the moments carrying information.

The demo can find keys, but what about an hour of security footage?

The dynamic scan is ideal for queries like “when did the person in the red jacket enter the room,” where the information density is low and most of the video is filler. Where the agentic approach might struggle, conversely, is complex analysis of overall action or subtle behavioral changes over time, where the model might not correctly guess which segments to skip and loses context.

Scanning error rates will determine real savings

The key metric now will be how often the model, in an effort to save tokens, skips a crucial frame and returns a hallucination instead of diligently scanning the rest of the track. If it's necessary to fall back to the old brute-force approach for critical tasks, the promised 66% cost savings will quickly evaporate.

Lilith's verdict

This shifts video from expensive and slow context to the level of smart database querying, where the agent only pays for what it actually reads.