BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

FrontierMath Tier 4

The hardest FrontierMath tier: problems that take expert mathematicians days. Run by Epoch AI.

As of 24 Oct 2026, GPT-6.1 Sol leads FrontierMath Tier 4 on BenchLeader with 100.0%, ahead of GPT-6 Astra at 97.6%, across 74 model configurations with a published result.

Published by
Epoch AI Benchmarking Hub
Category
Maths
Index weight
0.5
Models
74
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

Tier 4 is the hardest FrontierMath set: roughly fifty research problems that expert mathematicians estimate would take days of focused work. They are written to be beyond current systems and to test genuine mathematical reasoning rather than recall.

How it is scored

Percent of problems solved with an automatically checked final answer, run by Epoch AI under the same harness as the lower tiers.

What to keep in mind

With so few problems, a single solve moves the score by about two points, and most models score in the single digits. Treat it as a frontier signal, not a ranking of ordinary mathematical ability.

Top 15
  1. 1GPT-6.1 Sol (max)100.0%
  2. 2GPT-6 Astra (max)97.6%
  3. 3GPT-6 Astra (high)97.6%
  4. 4GPT-6 Astra (xhigh)97.6%
  5. 5GPT-6 Astra (medium)97.6%
  6. 6Claude Opus 5.5 (max)95.0%
  7. 7Claude Fable 5 (max)90.2%
  8. 8GPT-6 Sol (max)90.0%
  9. 9Claude Fable 5.1 (max)87.8%
  10. 10GPT-6 Astra (low)87.8%
  11. 11GPT-5.6 Sol (max)82.9%
  12. 12GPT-6 Astra (none)82.9%
  13. 13Claude Sonnet 5.5 (max)80.5%
  14. 14GPT-5.6 Sol Pro (max)80.5%
  15. 15Gemini 4 Argon (high)78.0%
74 of 74
#
1GPT-6.1 Solmax100.0%69.92026-09-29
2GPT-6 Astramax97.6%70.42026-09-03
3GPT-6 Astrahigh97.6%69.62026-09-03
4GPT-6 Astraxhigh97.6%68.92026-09-03
5GPT-6 Astramedium97.6%67.52026-09-03
6Claude Opus 5.5max95.0%72.62026-09-22
7Claude Fable 5max90.2%69.32026-06-09
8GPT-6 Solmax90.0%66.42026-09-22
9Claude Fable 5.1max87.8%70.52026-09-01
10GPT-6 Astralow87.8%65.92026-09-03
11GPT-5.6 Solmax82.9%68.02026-07-09
12GPT-6 Astrano reasoning82.9%59.02026-09-03
13Claude Sonnet 5.5max80.5%69.32026-09-28
14GPT-5.6 Sol Promax80.5%–2026-07-09
15Gemini 4 Argonhigh78.0%70.82026-09-30
16GPT-5.5 Proxhigh78.0%65.32026-04-23
17AI Co-Mathematician75.6%–2026-05-08
18Claude Opus 5max73.2%68.82026-07-24
19GPT-5.5xhigh72.5%66.12026-04-23
20Muse Wondermax70.7%67.52026-10-12
21GPT-5.6 Terramax70.7%64.42026-07-09
22GPT-5.6 Lunamax61.0%59.92026-07-09
23GPT-5.4 Proxhigh58.5%61.92026-03-05
24Claude Opus 4.8max56.1%63.32026-05-28
25GPT-6 Lunamax56.1%60.02026-09-22
26GPT-5.4xhigh49.0%64.72026-03-05
27Muse Spark 1.3max46.3%67.12026-09-02
28NewDeepSeek V4.1 Flash 1017max46.3%64.02026-10-17
29Claude Haiku 5.5max46.3%60.52026-10-07
30Qwen3.8 Maxxhigh46.3%57.12026-08-02
31GPT-5.2 Proxhigh46.0%–2025-12-11
32Muse Spark 1.3xhigh41.5%64.92026-09-02
33Kimi K3max39.0%64.72026-07-16
34Gemini 3.7 Flashhigh36.6%63.52026-08-13
35Qwen3.7max34.1%60.62026-05-19
36Grok 4.6xhigh31.7%62.52026-08-12
37Claude Opus 4.7max31.7%62.42026-04-16
38NewMuse Spark Open 816B A42Bxhigh31.7%58.62026-10-15
39GPT-5.2xhigh31.7%60.72025-12-11
40GLM 5.3max29.3%63.52026-08-14
41GLM 5.2max29.3%62.32026-06-16
42Claude Sonnet 5max29.3%58.42026-06-30
43Gemini 3.5 Flashhigh26.8%63.12026-05-19
44DeepSeek V4.1 Flashmax26.8%61.22026-09-09
45DeepSeek V4 Promax26.8%60.82026-08-13
46Gemini 3.1 Pro26.8%60.62026-02-19
47Claude Opus 4.6max26.8%57.42026-02-05
48Mistral Large 4high26.8%49.82026-10-06
49Kimi K2.625.6%55.42026-04-20
50DeepSeek V4 Flashmax24.4%59.42026-07-31
51Grok 4.5high24.4%59.32026-07-08
52Gemini 3.8 Flashhigh21.9%64.42026-09-02
53Gemini 3.6 Flashhigh21.9%62.02026-07-21
54GPT-5high21.9%57.92025-08-07
55GPT-5 Prohigh19.5%–2025-10-07
56Grok 4.7xhigh17.1%61.62026-09-21
57Grok 4.20thinking17.1%58.62026-02-17
58Gemini 3 Flash17.1%57.02025-12-17
59GLM 5.3 Flashmax17.1%54.22026-08-20
60Inkling Smallxhigh17.1%47.62026-07-15

More maths benchmarks · All LLM benchmarks

Cite as: BenchLeader, “FrontierMath Tier 4 leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/frontiermath_t4.

FrontierMath Tier 4: questions

What does FrontierMath Tier 4 measure?
Tier 4 is the hardest FrontierMath set: roughly fifty research problems that expert mathematicians estimate would take days of focused work. They are written to be beyond current systems and to test genuine mathematical reasoning rather than recall. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads FrontierMath Tier 4?
GPT-6.1 Sol leads FrontierMath Tier 4 with 100.0% as of 24 Oct 2026, ahead of GPT-6 Astra at 97.6%.
How many models have FrontierMath Tier 4 results?
74 model configurations have a FrontierMath Tier 4 result on BenchLeader, all taken from Epoch AI Benchmarking Hub.
Who runs FrontierMath Tier 4 and how often is it updated?
FrontierMath Tier 4 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 Tier 4 count toward the BenchLeader Index?
Yes. FrontierMath Tier 4 contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.