BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

FrontierMath Tiers 1–3

Unpublished research-level maths problems, tiers 1–3 (v2 private set). Run by Epoch AI.

As of 24 Oct 2026, GPT-6 Astra leads FrontierMath Tiers 1–3 on BenchLeader with 93.7%, ahead of GPT-6.1 Sol at 93.7%, across 118 model configurations with a published result.

Published by
Epoch AI Benchmarking Hub
Category
Maths
Index weight
1.0
Models
118
Data as of
24 Oct 2026
Newest result
17 Oct 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

What the test looks like

FrontierMath is a private set of research-level mathematics problems written by professional mathematicians, each with a definite, automatically checkable answer. Tiers 1 to 3 range from hard undergraduate to problems that take a specialist hours. The problems are unpublished, so they cannot appear in training data.

How it is scored

Percent of problems solved, verified by running the model's final answer through an automatic checker. Epoch AI runs the models itself, usually with a code tool available, and reports the mean of several runs.

What to keep in mind

Scores are low and noisy: a handful of problems separates models, and the confidence intervals Epoch publishes are wide. OpenAI funded the benchmark's creation and holds access to most of the problems, which Epoch discloses.

Top 15
  1. 1GPT-6 Astra (max)93.7%
  2. 2GPT-6.1 Sol (max)93.7%
  3. 3Claude Opus 5.5 (max)91.2%
  4. 4Claude Fable 5.1 (max)90.2%
  5. 5GPT-6 Sol (max)89.8%
  6. 6GPT-5.6 Sol (max)89.1%
  7. 7Claude Sonnet 5.5 (max)88.8%
  8. 8GPT-5.5 Pro (xhigh)87.7%
  9. 9Claude Fable 5 (max)87.0%
  10. 10Gemini 4 Argon (high)86.0%
  11. 11GPT-5.6 Terra (max)86.0%
  12. 12Claude Opus 5 (max)85.6%
  13. 13GPT-5.5 (xhigh)85.3%
  14. 14GPT-5.4 Pro (xhigh)82.5%
  15. 15Muse Wonder (max)82.1%
118 of 118
#
1GPT-6 Astramax93.7%70.42026-09-03
2GPT-6.1 Solmax93.7%69.92026-09-29
3Claude Opus 5.5max91.2%72.62026-09-22
4Claude Fable 5.1max90.2%70.52026-09-01
5GPT-6 Solmax89.8%66.42026-09-22
6GPT-5.6 Solmax89.1%68.02026-07-09
7Claude Sonnet 5.5max88.8%69.32026-09-28
8GPT-5.5 Proxhigh87.7%65.32026-04-23
9Claude Fable 5max87.0%69.32026-06-09
10Gemini 4 Argonhigh86.0%70.82026-09-30
11GPT-5.6 Terramax86.0%64.42026-07-09
12Claude Opus 5max85.6%68.82026-07-24
13GPT-5.5xhigh85.3%66.12026-04-23
14GPT-5.4 Proxhigh82.5%61.92026-03-05
15Muse Wondermax82.1%67.52026-10-12
16GPT-5.6 Lunamax82.1%59.92026-07-09
17Claude Opus 4.8max80.0%63.32026-05-28
18GPT-6 Lunamax79.0%60.02026-09-22
19GPT-5.4xhigh78.6%64.72026-03-05
20Claude Haiku 5.5max75.1%60.52026-10-07
21Qwen3.8 Maxxhigh74.7%57.12026-08-02
22Muse Spark 1.3xhigh74.4%64.92026-09-02
23Muse Spark 1.3max74.0%67.12026-09-02
24GPT-5.2 Proxhigh74.0%–2025-12-11
25Kimi K3max72.2%64.72026-07-16
26Gemini 3.7 Flashhigh71.6%63.52026-08-13
27NewDeepSeek V4.1 Flash 1017max71.2%64.02026-10-17
28Claude Opus 4.7max70.2%62.42026-04-16
29NewMuse Spark Open 816B A42Bxhigh69.8%58.62026-10-15
30GLM 5.3max68.8%63.52026-08-14
31Gemini 3.8 Flashhigh68.4%64.42026-09-02
32GPT-5.2xhigh67.4%60.72025-12-11
33DeepSeek V4.1 Flashmax67.4%61.22026-09-09
34Grok 4.6xhigh66.0%62.52026-08-12
35Claude Opus 4.6max66.0%57.42026-02-05
36Claude Sonnet 5max65.6%58.42026-06-30
37DeepSeek V4 Promax64.6%60.82026-08-13
38Qwen3.7max64.6%60.62026-05-19
39Gemini 3.5 Flashhigh62.8%63.12026-05-19
40Gemini 3.1 Pro59.6%60.62026-02-19
41GLM 5.2max59.2%62.32026-06-16
42Gemini 3.6 Flashhigh59.0%62.02026-07-21
43Mistral Large 4high58.6%49.82026-10-06
44DeepSeek V4 Flashmax57.5%59.42026-07-31
45Grok 4.5high57.2%59.32026-07-08
46Kimi K2.657.2%55.42026-04-20
47GLM 5.3 Flashmax55.8%54.22026-08-20
48GPT-5 Prohigh55.8%–2025-10-07
49GPT-5high55.4%57.92025-08-07
50GLM 5.2low54.7%–2026-06-16
51Kimi K2.7 Code54.0%55.82026-06-12
52Grok 4.7xhigh53.0%61.62026-09-21
53Gemini 3 Flash51.2%57.02025-12-17
54GPT-5.4 Minixhigh51.2%56.32026-03-17
55GPT-5 minihigh46.7%54.12025-08-07
56Inkling Smallxhigh46.3%47.62026-07-15
57Grok 4.20thinking44.9%58.62026-02-17
58GPT-5.4 Nanohigh44.9%48.42026-03-17
59Grok 4.3high42.8%57.02026-04-17
60GLM 5.2no reasoning42.5%48.02026-06-16

More maths benchmarks · All LLM benchmarks

Cite as: BenchLeader, “FrontierMath Tiers 1–3 leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/frontiermath_t1_3.

FrontierMath Tiers 1–3: questions

What does FrontierMath Tiers 1–3 measure?
FrontierMath is a private set of research-level mathematics problems written by professional mathematicians, each with a definite, automatically checkable answer. Tiers 1 to 3 range from hard undergraduate to problems that take a specialist hours. The problems are unpublished, so they cannot appear in training data. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads FrontierMath Tiers 1–3?
GPT-6 Astra leads FrontierMath Tiers 1–3 with 93.7% as of 24 Oct 2026, ahead of GPT-6.1 Sol at 93.7%.
How many models have FrontierMath Tiers 1–3 results?
118 model configurations have a FrontierMath Tiers 1–3 result on BenchLeader, all taken from Epoch AI Benchmarking Hub.
Who runs FrontierMath Tiers 1–3 and how often is it updated?
FrontierMath Tiers 1–3 is published by Epoch AI Benchmarking Hub. BenchLeader re-reads the published results once a day and records the date each result was published.
Does FrontierMath Tiers 1–3 count toward the BenchLeader Index?
Yes. FrontierMath Tiers 1–3 contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.