What capabilities does FrontierMath test?

FrontierMath evaluates math, scientific reasoning.

What is the current top score on FrontierMath?

The top reported score is 97.6% by GPT-6 Astra, across 69 models reporting (40 from frontier labs).

FrontierMath is available under Closed.

Saturated

Unpublished collection of research-level mathematics problems written by professional mathematicians, designed to be the hardest open math benchmark.

Publisher: Epoch AI
Capabilities: Math Scientific Reasoning
Domain: math
Format: Custom
Size: 300 tasks
License: Closed
Published: Nov 2024
Updates: Monthly
Notable for: The reference frontier-math benchmark — problems are kept private to prevent contamination, and even top reasoning models scored under 10% at release.
Canonical: epoch.ai/frontiermath
Official leaderboard: epoch.ai/frontiermath

Cite

Notes

Only stored in your browser.

Attribution

Top score 97.6% by GPT-6 Astra - 69 models reporting (40 frontier)

FrontierMathBar chart with 21 bars. Highest value: GPT-6 Astra at 97.6.

21 models

epoch.ai

preprint · 2024

Epoch AI benchmark of hundreds of original research-level math problems authored by professional mathematicians, with auto-verifiable answers.

preprint · 2024

Epoch AI benchmark of hundreds of original research-level math problems authored by professional mathematicians, with auto-verifiable answers.

What is FrontierMath?: Unpublished collection of research-level mathematics problems written by professional mathematicians, designed to be the hardest open math benchmark.
What capabilities does FrontierMath test?: FrontierMath evaluates math, scientific reasoning.
What is the current top score on FrontierMath?: The top reported score is 97.6% by GPT-6 Astra, across 69 models reporting (40 from frontier labs).
What license is FrontierMath under?: FrontierMath is available under Closed.