It's both a token cost and a performance cost; there's only so much that documentation can do, compared to a ton of RL on top of millions of lines of examples. The space is a continuum, but the more you stray from the trained path the higher the cost you pay.