Caveman doesn't and cannot change the tokenizer, so the relative token count differences by input category will remain unchanged.