I don't think that's correct. Increasing parameter count increases capabilities in all domains per the scaling laws. Models are larger than they were years ago, so capabilities in all domains must necessarily be better. This doesn't even account for better training data, which has also much improved.