Have you guys thought about using computationally-cheap, old-school NLP methods (such as Flesch-Kincaid readability, or heuristic methods for counting # of subordinate clauses) to determine whether it's worth paying for a more expensive model on a per-query basis?