When a Hard AI Exam Starts Looking Easy, the Measurement Problem Gets Real

A benchmark built to stress frontier models is showing faster progress than expected, and that may tell us as much about the test itself as about the machines taking it.