Math

FrontierMath v2 (Tier 4) leaderboard

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

As of 2026-10-07, the highest measured score on FrontierMath v2 (Tier 4) is 97.6% by GPT-6 Astra. 12 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	GPT-6 Astra	97.6%	measured
2	GPT-5.6 Sol	83.0%	measured
3	GPT-5.6 Terra	68.3%	measured
4	GPT-5.6 Luna	58.5%	measured
5	GPT-5.5 Pro	39.6%	measured
6	GPT-5.4 Pro	37.5%	measured
7	GPT-5.5	35.4%	measured
8	Claude Opus 4.7 (Adaptive)	35.2%	estimated ± 5.0 pp, low confidence
9	Claude Opus 4.8	31.3%	measured
10	GLM-5.2	27.8%	estimated ± 3.7 pp, low confidence
11	GPT-5.4	27.1%	measured
12	Claude Opus 4.7	22.9%	measured
13	Claude Opus 4.6	22.9%	measured
14	GPT-5.2	18.8%	measured
15	Gemini 3 Pro	18.8%	measured
16	Gemini 3.1 Pro	16.7%	measured
17	Muse Spark	14.6%	measured
18	Gemini 3.5 Flash	14.6%	measured
19	Kimi K2.6	14.6%	measured
20	GLM-5.1	12.5%	measured
21	GPT-5.1	12.5%	measured
22	Qwen3.6 Plus	8.3%	measured
23	Claude Sonnet 4.6	8.3%	measured
24	GPT-5.4 nano	6.3%	measured
25	o4-mini (high)	6.3%	measured
26	Kimi K2.5	4.2%	measured
27	Claude Opus 4.5	4.2%	measured
28	Claude Sonnet 4.5	4.2%	measured
29	Gemini 2.5 Flash	4.2%	measured
30	Gemini 2.5 Pro	4.2%	measured
31	Gemini 3 Flash	4.2%	measured
32	Qwen 3.6 Max (preview)	4.2%	measured
33	GLM-4.6	2.1%	measured
34	DeepSeek V3.2	2.1%	measured
35	GLM-5	2.1%	measured
36	Claude Haiku 4.5	2.1%	measured
37	Grok 4	2.1%	measured
38	o3	2.1%	measured
39	Qwen3.5 Plus	2.1%	measured
40	GPT-5.4 mini	2.1%	measured
41	o1	1.2%	estimated ± 4.5 pp, high confidence
42	GPT-4.1 mini	1.2%	estimated ± 4.5 pp, high confidence
43	DeepSeek V3	1.2%	estimated ± 4.5 pp, medium confidence
44	GPT-4.1 nano	1.2%	estimated ± 4.5 pp, medium confidence
45	GPT-4o	1.2%	estimated ± 4.5 pp, medium confidence
46	Llama 4 Maverick	1.2%	estimated ± 4.5 pp, medium confidence
47	Llama 4 Scout	1.2%	estimated ± 4.5 pp, medium confidence
48	Qwen3.5 397B	0.1%	estimated ± 3.7 pp, low confidence
49	Claude 3.5 Sonnet	0.0%	measured
50	GLM-4.7	0.0%	measured
51	GPT-4.1	0.0%	measured
52	Grok 3 [Beta]	0.0%	measured
53	Kimi K2	0.0%	measured
54	Qwen3 235B 2507 (Reasoning)	0.0%	measured
55	Qwen3.5 Flash	0.0%	measured
56	Qwen3.6-27B	0.0%	estimated ± 3.7 pp, low confidence
57	Qwen3.6-35B-A3B	0.0%	estimated ± 3.7 pp, low confidence

FrontierMath v2 (Tier 4) leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General