FrontierMath-Tiers-1-3-v2-Private

FrontierMath

Rank #10 of 51 · 31 models scored · top score 87.72

Research-level novel math, far beyond competition

Strongest on discrimination, weakest on saturation headroom.

consistent harnessheld-out test set

Headroom to ceiling

Score distribution

State of the art over time

Integrity breakdown

Weighted composite. Each signal is computed from the source-backed leaderboard below.

Discrimination30%100

Does it still separate models?

Saturation headroom30%94

How far from ceiling / clustered at the top?

Contamination resistance15%100

Public vs held-out; training-leak risk.

Harness comparability15%100

Apples-to-apples, or mixed / vendor-optimized?

Freshness10%100

How old is the benchmark?

Leaderboard

#	Model	Score	Evidence	Measured
1	GPT-5.5 Pro	87.72	reproduced	2026-04-23
2	Claude Fable 5	87.02	reproduced	2026-06-09
3	GPT-5.5	85.26	reproduced	2026-04-23
4	GPT-5.4 Pro	82.46	reproduced	2026-03-05
5	Claude Opus 4.8	80	reproduced	2026-05-28
6	GPT-5.4	78.6	reproduced	2026-03-05
7	GPT-5.2 Pro	74	reproduced	2025-12-11
8	Claude Opus 4.7	70.18	reproduced	2026-04-16
9	GPT-5.2	67.4	reproduced	2025-12-11
10	Claude Opus 4.6	65.96	reproduced	2026-02-05
11	Qwen3.7-Max	64.56	reproduced	2026-05-19
12	Gemini 3.5 Flash	62.81	reproduced	2026-05-19
13	Gemini 3.1 Pro	59.65	reproduced	2026-02-19
14	GLM-5.2	59.21	reproduced	2026-06-16
15	Kimi K2.6	57.19	reproduced	2026-04-20
16	GPT-5 Pro	55.79	reproduced	2025-10-07
17	GPT-5	55.44	reproduced	2025-08-07
18	Kimi K2.7 Code	54.04	reproduced	2026-06-12
19	Gemini 3 Flash	51.23	reproduced	2025-12-17
20	GPT-5.4 Mini	51.23	reproduced	2026-03-17
21	GPT-5 mini	46.67	reproduced	2025-08-07
22	DeepSeek-V4-Pro	45.26	reproduced	2026-04-24
23	GPT-5.4 Nano	44.91	reproduced	2026-03-17
24	Grok 4.3 Beta	42.81	reproduced	2026-04-17
25	o4-mini	36.14	reproduced	2025-04-16
26	Claude Opus 4.5	34.39	reproduced	2025-11-24
27	Gemini 2.5 Pro (Jun 2025)	24.56	reproduced	2025-06-05
28	Claude Sonnet 4.5	23.86	reproduced	2025-09-29
29	GPT-5 nano	20	reproduced	2025-08-07
30	o3-mini	18.6	reproduced	2025-01-31
31	Claude Opus 4.1	12.63	reproduced	2025-08-05

At a glance

Models scored: 31
Top model: GPT-5.5 Pro
Top score: 87.72
Score spread: 75.09
Headroom: 12.28

Benchmark context

Task type: math-research
Metric: Exact-match auto-verified (SymPy/numeric)
Score ceiling: 100
Human baseline: No human solves the set; hours-days per problem
Test-set privacy: private
Construction: 60+ mathematicians; v2 ~338 Tier1-3 + Tier4; auto-checkable answers
Key assumptions: Novel held-out + auto-verify = contamination-proof, measures discovery
Limitations: Excludes proofs; tiny; correct answer can hide invalid method; access controversy
Contamination history: OpenAI (funder) had access to ~28/48 Tier-4 under verbal no-train; undisclosed funding scandal 2025
Harness notes: Single-shot + Python + heavy test-time compute; best ~30-45% Tier1-3
Authors: Epoch AI (OpenAI-funded)
Released: 2024-11 (v2 2026-06)
License: Proprietary (private held by Epoch)
Source: epoch.ai/frontiermath

Provenance

31 source records · scores from epoch_benchmarks. Every fact on this page is evidence-linked.

Follow the record

Integrity neighbors

Top models here

Head-to-head

Integrity breakdown

Weighted composite. Each signal is computed from the source-backed leaderboard below.

Discrimination30%100

Does it still separate models?

Saturation headroom30%94

How far from ceiling / clustered at the top?

Contamination resistance15%100

Public vs held-out; training-leak risk.

Harness comparability15%100

Apples-to-apples, or mixed / vendor-optimized?

Freshness10%100

How old is the benchmark?

Leaderboard

#	Model	Score	Evidence	Measured
1	GPT-5.5 Pro	87.72	reproduced	2026-04-23
2	Claude Fable 5	87.02	reproduced	2026-06-09
3	GPT-5.5	85.26	reproduced	2026-04-23
4	GPT-5.4 Pro	82.46	reproduced	2026-03-05
5	Claude Opus 4.8	80	reproduced	2026-05-28
6	GPT-5.4	78.6	reproduced	2026-03-05
7	GPT-5.2 Pro	74	reproduced	2025-12-11
8	Claude Opus 4.7	70.18	reproduced	2026-04-16
9	GPT-5.2	67.4	reproduced	2025-12-11
10	Claude Opus 4.6	65.96	reproduced	2026-02-05
11	Qwen3.7-Max	64.56	reproduced	2026-05-19
12	Gemini 3.5 Flash	62.81	reproduced	2026-05-19
13	Gemini 3.1 Pro	59.65	reproduced	2026-02-19
14	GLM-5.2	59.21	reproduced	2026-06-16
15	Kimi K2.6	57.19	reproduced	2026-04-20
16	GPT-5 Pro	55.79	reproduced	2025-10-07
17	GPT-5	55.44	reproduced	2025-08-07
18	Kimi K2.7 Code	54.04	reproduced	2026-06-12
19	Gemini 3 Flash	51.23	reproduced	2025-12-17
20	GPT-5.4 Mini	51.23	reproduced	2026-03-17
21	GPT-5 mini	46.67	reproduced	2025-08-07
22	DeepSeek-V4-Pro	45.26	reproduced	2026-04-24
23	GPT-5.4 Nano	44.91	reproduced	2026-03-17
24	Grok 4.3 Beta	42.81	reproduced	2026-04-17
25	o4-mini	36.14	reproduced	2025-04-16
26	Claude Opus 4.5	34.39	reproduced	2025-11-24
27	Gemini 2.5 Pro (Jun 2025)	24.56	reproduced	2025-06-05
28	Claude Sonnet 4.5	23.86	reproduced	2025-09-29
29	GPT-5 nano	20	reproduced	2025-08-07
30	o3-mini	18.6	reproduced	2025-01-31
31	Claude Opus 4.1	12.63	reproduced	2025-08-05

Benchmark context

Task type

math-research

Metric

Exact-match auto-verified (SymPy/numeric)

Score ceiling

100

Human baseline

No human solves the set; hours-days per problem

Test-set privacy

private

Construction

60+ mathematicians; v2 ~338 Tier1-3 + Tier4; auto-checkable answers

Key assumptions

Novel held-out + auto-verify = contamination-proof, measures discovery

Limitations

Excludes proofs; tiny; correct answer can hide invalid method; access controversy

Contamination history

OpenAI (funder) had access to ~28/48 Tier-4 under verbal no-train; undisclosed funding scandal 2025

Harness notes

Single-shot + Python + heavy test-time compute; best ~30-45% Tier1-3

Authors

Epoch AI (OpenAI-funded)

Released

2024-11 (v2 2026-06)

License

Proprietary (private held by Epoch)

Source

epoch.ai/frontiermath