benchmarks.site

Public leaderboard · Math

FrontierMath

Hundreds of brand-new, unpublished math problems written by professional mathematicians, ranging from hard competition level up to problems that take experts days. Because the problems are secret, it is a strong check of real mathematical ability rather than memorization.

Leaderboard

Top 3 right now

  1. 1stClaude Fable 5.2 (max)95.1
  2. 2ndGPT-6.1 Sol (max)93.7
  3. 3rdGPT-6 Astra (max)93.7

What it measures

Unpublished, expert-written math problems in four difficulty tiers up to research level, with answers checked automatically, scored by accuracy.

How to read the numbers

Share of Tiers 1-3 problems solved correctly (%) — higher is better.

Results

25 rows · higher is better (Tiers 1-3 (%))CSV
1Claude Fable 5.2 (max)Anthropic95.1100.02026-11-09
2GPT-6.1 Sol (max)OpenAI93.7100.02026-09-29
3GPT-6 Astra (max)OpenAI93.797.62026-09-03
4Claude Opus 5.5 (max)Anthropic91.295.02026-09-22
5GPT-6.1 Terra (max)OpenAI90.992.72026-10-26
6Gemini 4 Carbon (high)Google DeepMind90.595.12026-11-10
7Claude Fable 5.1 (max)Anthropic90.287.82026-09-01
8GPT-6 Sol (max)OpenAI89.890.02026-09-22
9GPT-5.6 Sol (max)OpenAI89.182.92026-07-09
10Claude Sonnet 5.5 (max)Anthropic88.880.52026-09-28
11GPT-5.5 Pro (xhigh)OpenAI87.778.02026-04-23
12Claude Fable 5 (max)Anthropic87.090.22026-06-09
13Gemini 4 Argon (high)Google DeepMind86.078.02026-10-13
14GPT-5.6 Terra (max)OpenAI86.070.72026-07-09
15Claude Opus 5 (max)Anthropic85.673.22026-07-24
16GPT-5.5 (xhigh)OpenAI85.372.52026-04-23
17GPT-5.4 Pro (xhigh)OpenAI82.558.52026-03-05
18Muse Wonder (max)Meta AI82.170.72026-10-12
19GPT-5.6 Luna (max)OpenAI82.161.02026-07-09
20Claude Opus 4.8Anthropic80.056.12026-05-28
21GPT-6 Luna (max)OpenAI78.956.12026-09-22
22GPT-5.4 (xhigh)OpenAI78.649.02026-03-05
23Kimi K3.1 Pro (Together, max)Moonshot78.265.92026-10-23
24Claude Haiku 5.5 (max)Anthropic75.146.32026-10-07
25Qwen3.8 Max (xhigh)Alibaba74.746.32026-08-02

Note: Epoch AI runs on the private v2 problem sets (frontiermath_tiers_1_3_v2.csv and frontiermath_tier_4_v2.csv), converted to % (1 decimal). Tier 4 is the research-level tier; blank = not evaluated on Tier 4 v2. OpenAI funded the benchmark and has access to much of the problem set.

Source: Epoch AI (Benchmarking Hub) · epoch.ai · results as of 2026-11-17 · retrieved 2026-11-17 · license: CC-BY-4.0

Markdown · More Math