Benchmark · Reasoning and knowledge

FrontierMath Tier 4

The hardest FrontierMath problems, an expansion set of 43 in version 2, each taking a specialist researcher days. Same method as FrontierMath's Tiers 1-3, run by Epoch AI on the private set.

Score
% of problems solved, higher is better
Run by
Epoch AI
Status
Current, newest result 19 November 2026
Updated
as models are released
Ceiling
The best result is within five points of a perfect score, so the test no longer tells the top models apart
Data
the published file, CC BY 4.0, read 20 November 2026
Method
epoch.ai/benchmarks/frontiermath-tier-4-v2

Results: Tier 4 v2 (2026-06-12)

32 models we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1Claude Fable 5.2 A100.0%Independenteffort max2026-11-06source
1GPT-6.2 Astra AA100.0%Independenteffort max2026-10-27source
1GPT-6.1 Sol AA100.0%Independenteffort max2026-09-29source
4GPT-6 Astra A97.6%Independenteffort high2026-08-30source
5Gemini 4 Neon B97.6%Independenteffort xhigh2026-11-19source
6Gemini 4 Carbon B95.1%Independenteffort high2026-11-02source
7Claude Opus 5.5 A95.0%Independenteffort max2026-09-22source
8GPT-6.1 Terra AA92.7%Independenteffort max2026-10-26source
9GPT-6 Sol A90.0%Independenteffort max2026-09-22source
10Gemini 4 Hydrogen B87.8%Independenteffort xhigh2026-11-13source
10Claude Fable 5.1 A87.8%Independenteffort max2026-09-01source
12Claude Sonnet 5.5 A80.5%Independenteffort max2026-09-29source
13Gemini 4 Argon B78.0%Independenteffort high2026-10-13source
14Muse Wonder D70.7%Independenteffort max2026-10-12source
15Kimi K3.1 Pro D65.8%Independenteffort max2026-11-06source
16GPT-6 Luna A56.1%Independenteffort max2026-09-22source
17DeepSeek V4.1 Flash C46.3%Independenteffort max2026-10-19source
17Claude Haiku 5.5 A46.3%Independenteffort max2026-10-07source
17Muse Spark 1.3 D46.3%Independenteffort max2026-09-18source
20Kimi K3 B39.0%Independenteffort max2026-07-17source
21Muse Spark Open C31.7%Independenteffort xhigh2026-10-15source
22GLM-5.3 B29.3%Independenteffort max2026-08-25source
23Gemini 3.1 Pro (preview) AA26.8%Independent2026-06-11source
24DeepSeek-V4-Flash-0731 C24.4%Independenteffort max2026-08-02source
25Gemini 3.8 Flash AA21.9%Independenteffort high2026-09-02source
26Grok 4.7 BB17.1%Independenteffort xhigh2026-09-22source
26GLM-5.3-Flash B17.1%Independenteffort max2026-08-27source
26Inkling-Small B17.1%Independenteffort xhigh2026-08-14source
29Grok 4.3 BB14.6%Independenteffort high2026-06-17source
30Inkling B4.9%Independenteffort xhigh2026-08-06source
31DeepSeek V4 Pro D2.4%Independenteffort max2026-06-17source
32Gemini 3.5 Flash-Lite AA0.0%Independenteffort high2026-08-02source
Every setting, 39 rows
ModelAs the leaderboard names itScoreWho ran itSettingDate
Claude Fable 5.2claude-fable-5-2100.0%Independenteffort max2026-11-06source
GPT-6.2 Astragpt-6.2-astra100.0%Independenteffort max2026-10-27source
GPT-6.1 Solgpt-6.1-sol100.0%Independenteffort max2026-09-29source
GPT-6 Astragpt-6-astra97.6%Independenteffort high2026-08-30source
GPT-6 Astragpt-6-astra97.6%Independenteffort max2026-08-30source
GPT-6 Astragpt-6-astra97.6%Independenteffort xhigh2026-08-30source
Gemini 4 Neongemini-4-neon97.6%Independenteffort xhigh2026-11-19source
GPT-6 Astragpt-6-astra97.6%Independenteffort medium2026-08-30source
Gemini 4 Carbongemini-4-carbon95.1%Independenteffort high2026-11-02source
Claude Opus 5.5claude-opus-5-595.0%Independenteffort max2026-09-22source
GPT-6.1 Terragpt-6.1-terra92.7%Independenteffort max2026-10-26source
GPT-6 Solgpt-6-sol90.0%Independenteffort max2026-09-22source
Gemini 4 Hydrogengemini-4-hydrogen87.8%Independenteffort xhigh2026-11-13source
Claude Fable 5.1claude-fable-5-187.8%Independenteffort max2026-09-01source
GPT-6 Astragpt-6-astra87.8%Independenteffort low2026-08-30source
GPT-6 Astragpt-6-astra82.9%Independenteffort none2026-08-30source
Claude Sonnet 5.5claude-sonnet-5-580.5%Independenteffort max2026-09-29source
Gemini 4 Argongemini-4-argon78.0%Independenteffort high2026-10-13source
Muse Wondermuse-wonder-1.070.7%Independenteffort max2026-10-12source
Kimi K3.1 Prokimi-k3.1-pro-together65.8%Independenteffort max2026-11-06source
GPT-6 Lunagpt-6-luna56.1%Independenteffort max2026-09-22source
DeepSeek V4.1 Flashdeepseek-v4.1-flash-101746.3%Independenteffort max2026-10-19source
Claude Haiku 5.5claude-haiku-5-546.3%Independenteffort max2026-10-07source
Muse Spark 1.3muse-spark-1.346.3%Independenteffort max2026-09-18source
Muse Spark 1.3muse-spark-1.341.5%Independenteffort xhigh2026-09-16source
Kimi K3kimi-k339.0%Independenteffort max2026-07-17source
Muse Spark OpenMuse-Spark-Open-816B-A42B31.7%Independenteffort xhigh2026-10-15source
GLM-5.3glm-5.329.3%Independenteffort max2026-08-25source
DeepSeek V4.1 Flashdeepseek-v4.1-flash26.8%Independenteffort max2026-10-08source
Gemini 3.1 Pro (preview)gemini-3.1-pro-preview26.8%Independent2026-06-11source
DeepSeek-V4-Flash-0731deepseek-v4-flash-073124.4%Independenteffort max2026-08-02source
Gemini 3.8 Flashgemini-3.8-flash21.9%Independenteffort high2026-09-02source
Grok 4.7grok-4.717.1%Independenteffort xhigh2026-09-22source
GLM-5.3-Flashglm-5.3-flash17.1%Independenteffort max2026-08-27source
Inkling-Smallinkling-small17.1%Independenteffort xhigh2026-08-14source
Grok 4.3grok-4.314.6%Independenteffort high2026-06-17source
InklingInkling4.9%Independenteffort xhigh2026-08-06source
DeepSeek V4 Prodeepseek-v4-pro2.4%Independenteffort max2026-06-17source
Gemini 3.5 Flash-Litegemini-3.5-flash-lite0.0%Independenteffort high2026-08-02source

Results: Tier 4 (2025-07-01)

1 model we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1Gemini 3.1 Pro (preview) AA16.7%Independent2026-02-19source

A rank is among the models we list with a result, not a place on the whole leaderboard, and a model offered by several hosts is shown once, on its own listing. These results don't move our grades. Data: Epoch AI, Capabilities and benchmarking (CC BY 4.0), CC BY 4.0.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.