The named tests everyone quotes. Each measures something quite specific, and the names give no clue what. Knowing roughly what the common ones cover is the difference between reading a model announcement critically and taking the bar chart at face value.
Technical view
Broad knowledge and reasoning suites test multiple-choice academic questions across subjects; mathematical benchmarks range from grade-school word problems to competition level; coding benchmarks measure functional correctness of generated programs against test suites, with agentic variants requiring changes to real repositories; long-context suites test retrieval and aggregation across large inputs. Most saturate within a couple of years of release, which is why leaderboards churn and why comparisons across time are unreliable.
Related
Part of AI Terminology Index, section: Measuring Performance.