What Changed Is Who Decides What to Watch
The old approach was static: sample the entire video at a fixed frame rate, 1 FPS by default with developer adjustment through the API. On long-form video the problem is immediate — a 90-minute lecture at 1 FPS is 5,400 frames, which forces a choice between high token costs and downsampling tricks that routinely drop the details that mattered. The new approach hands the decision to the model. It combines its own reasoning with native video tools to dynamically search, scan and inspect target segments, selecting between visual frames, audio and transcript as appropriate. The model works out where to look for this particular question and then looks there, instead of watching everything before answering. The nature of the change is closer to retrieval than compression: what gets saved is not quality but the content that was never loaded in the first place.
The Numbers and Their Conditions
Google's figures: up to 66% lower analysis costs and up to 88% fewer tokens on standard video analysis benchmarks, with accuracy improving up to 7%. The gains concentrate on long-form video — from 10-minute how-to guides to 90-minute lectures and multi-hour recordings. The pricing documentation is more grounded: token usage varies with the content the model loads rather than full video length, typically yielding up to 88% fewer input tokens on long-form video, though counts depend on query complexity and dynamic sampling depth, which may exceed 1 FPS for segments needing detailed visual inspection. In other words, 88% is a ceiling rather than an average, and bills will float with question difficulty. Asking "what is this video about" and asking "what brand is the object that person is holding at minute 47" will not cost the same.
How to Enable It, and What to Watch For
Set the API configuration to agentic in Google AI Studio or the Gemini Enterprise Agent Platform. It is labeled a Generative AI Preview with no additional feature fee. The Gemini app is listed as coming soon, and Ask YouTube on the watch page is listed for the coming months. Community reaction split. Some welcomed the 88% token reduction as what finally makes long-form video analysis practical and affordable; others questioned the models' reliability and argued Google should fix core quality first. That disagreement actually points at the same test priority: dynamic sampling means the model may simply never see a given key frame, whereas static sampling is at least predictable. To evaluate it, build a set of cases where the answer appears in only one or two frames and see whether it lands on them. For teams building video applications, the practical implication is that the cost structure changed: budgets used to scale with video duration and now scale with query type. Running your typical queries to get an actual token distribution beats applying the 88% figure before you ship.