Toggle a few realistic setups below. This is a simplified illustration built to show the shape of the problem — not a live measurement of any real account, model, or conversation.
Model: a 200,000-token window (roughly the size of a smaller current model), with illustrative token weights — short instructions ≈150, long instructions ≈2,000, a few project files ≈5,000, many project files ≈90,000, each connected app ≈500. Real products differ: some route large knowledge bases through retrieval instead of loading everything, and real connector or instruction sizes vary. See the article above for the sourcing behind the underlying mechanism.