I mean whatever models I use (with Claude code) sub agents seem to use absurd amounts of tokens for trivial (or at least small) tasks.