MATH Level 5

Citation frequency 1/3Saturated
What does it measure?
Tests math problem solving on only the hardest tier (level 5) of MATH, a collection of high-school competition problems.
Who evaluates it, and how?
1,324 level-5 problems from the MATH test set. Grading uses the final answer, not the working, and a match is checked by string comparison, mathematical equivalence, or an AI judge.
Reading the score
Scores run from 0 to 100%. Higher is better. Recent models mostly score high, so the top models are hard to tell apart.
Top score98.1%
Models tested10
Scoring basisHigher is better
Last updated2025.10.30
Model release date
  1. 1
    OpenAI2025.08.07 · high
    98.1%
  2. 2
    OpenAI2025.08.07 · high
    97.8%
  3. 3
    Anthropic2025.09.29 · budget-32K
    97.7%
  4. 4
    Anthropic2025.10.15 · budget-32K
    96.4%
  5. 5
    OpenAI2025.08.07 · medium
    95.2%
  6. 6
    Anthropic2025.05.22
    85.0%
  7. 7
    Anthropic2025.05.22
    84.4%
  8. 8
    OpenAI2025.04.14
    83.0%
  9. 9
    Meta2025.04.05
    73.0%
  10. 10
    Meta2025.04.05
    62.3%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.