MMLU

Introduced in 2020, MMLU spans roughly 14,000 multiple-choice questions across 57 subjects from elementary mathematics to professional law and medicine. For several years it was the headline number in nearly every model announcement, with GPT-4's 2023 result marking the moment models approached expert-level scores.

Frontier models now cluster in the high 80s and low 90s percent, close enough to the ceiling — and to the test's own label-error rate — that MMLU no longer separates top models meaningfully. Harder successors like MMLU-Pro, with more options and more reasoning-dependent questions, were built to restore discrimination.

MMLU remains useful as a floor check for mid-tier and small models, where scores still spread widely, and as a historical axis for tracking how quickly cheap models absorb capabilities that once commanded frontier prices.