Benchmark saturation
When top models score so close to the maximum on a benchmark that it can no longer tell them apart.
A benchmark is useful while scores spread out. As models improve, the best cluster near the ceiling, where remaining errors are often mistakes in the questions themselves. GLUE, a language understanding benchmark from 2018, was saturated within about a year, its successor SuperGLUE soon after, and MMLU, HumanEval and GSM8K by 2024. Saturation, along with contamination, is why the field keeps building harder tests such as GPQA, Humanity's Last Exam and tasks measured by how long an agent can work unaided.