MATH

MATH: Measuring Mathematical Problem Solving

12,500 competition mathematics problems (7,500 train, 5,000 test) from AMC 10/12, AIME and similar contests across seven subjects and five difficulty levels, each with a full LaTeX step-by-step solution. Scored as exact match of the boxed final answer after normalization. It was the primary hard-math benchmark until reasoning models pushed it past 95%; the hardest Level 5 subset and the 500-problem MATH-500 split remain in use.

98.1% · human 90%
o3-high
self-reported
2021-03-05 → 2025-04-16: 5.2% → 98.1%
Released
2021-03
Maintainer
Dan Hendrycks et al. (UC Berkeley)
Status
saturated
Contamination
high
Metric
accuracy (percent, ↑)
Tasks
5,000
Domains
math reasoning
human
90% three-time IMO gold medalist (single participant; a CS PhD student scored about 40%)

Notes. Answer-equivalence checking differs across harnesses (string match vs sympy vs LLM grader) and can move scores by several points. The human numbers in the paper are anecdotal single-person measurements.

Full ledger

SystemDeveloperScoreDateSourceConditions
o3-highOpenAI98.1%self-reportedsplit: MATH-500 shots: 0
simple-evals README; o4-mini-high 98.2 on the same table. Epoch's Level-5 run of gpt-5 (high) reached 98.1 in October 2025.
o1-2024-12-17 (medium)OpenAI94.4%independentsplit: Level 5
Epoch AI Benchmarking Hub run of 2025-01-27 on the 1,324 Level-5 test problems (data export).
gpt-4-turbo-2024-04-09OpenAI73.4%self-reportedsplit: MATH-500 shots: 0
simple-evals README; the MATH column is the 500-problem subset, zero-shot CoT.
GPT-3 175B (few-shot)OpenAI5.2%paper
Table 2 of the MATH paper, overall accuracy (175B*, few-shot). The paper's human anchors: CS PhD student about 40%, three-time IMO gold medalist 90%.