This isn't surprising but it's good to have reporting to point to. The only way a general purpose LLM like these will match the performance of a bespoke classifier is if it has access to such a model.