---
title: "FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI"
authors:
  - "Elliot Glazer"
  - "Ege Erdil"
  - "Tamay Besiroglu"
  - "Diego Franco"
  - "Jean-Stanislas Denain"
  - "Anson Ho"
  - "Matthew Barnett"
  - "Robert Sandler"
  - "Epoch AI collaborators"
year: 2024
venue: "arXiv preprint"
arxiv: "2411.04872"
doi: null
source: "https://arxiv.org/abs/2411.04872"
topics:
  - "mathematical-reasoning"
seed_rank: 1226
seed_batch: "alpoge-math-2026-08-24"
reviewed: "2026-08-24"
pool: "agents"
relevance_score: 9
lineage: ai-mathematical-reasoning
cites:
  []
see:
  - "1220-solving-quantitative-reasoning-problems-with-language-models"
---

# FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI

## One-sentence takeaway

Epoch AI's private, expert-authored research-grade math benchmark: leading 2024 models (o1-preview, GPT-4o, Claude 3.5, Gemini 1.5 Pro) all below 2% — the anti-GSM8K.

## Why it matters here

User named FrontierMath as the benchmark paper to mint. It is the eval the later IMO/o-series events are measured against, not a +2% MATH paper.

## Problem

MATH and GSM8K saturated; contamination is rife. Need unpublished, specialist problems that take humans hours or days.

## Design

Hundreds of new problems across number theory, algebraic geometry, etc., peer-reviewed, held private. Tiers 1-4 plus a later Open Problems track.

## Evidence

arXiv:2411.04872 (Nov 2024). epoch.ai/frontiermath. Commissioned in part by OpenAI (disclosed).

## Limitations

Not a solve paper. Private labels. OpenAI ownership of a slice is a contamination footnote. Putnam-Bench / MiniF2F are benches, not minted.

## Implications for Broadside

The eval card for the AI-math shelf. See o1 1228 and IMO gold 1229/1230 as the systems that later moved the needle.

## Bottom line

Mint FrontierMath. Skip Putnam-Bench and MiniF2F as standalone benches.

## Links

- arXiv: [2411.04872](https://arxiv.org/abs/2411.04872)
- https://epoch.ai/frontiermath