Baselines vary with rater expertise and time limits: PhD raters on GPQA are not Mechanical Turk on MMLU. Crossing a human baseline on a test does not mean the model matches humans on the job. It does tell you the exam is no longer a ceiling for that population.