still quite verbose at 140m output tokens, but this is on max thinking. high should do better.