I’ve noticed this with model upgrades too: sometimes they improve the “thinking” but lose the tightness of the writing.