Open-Weight Models Are Now Good Enough for 80% of Tasks