Math

FrontierMath (legacy) leaderboard

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

As of 2026-10-07, the highest measured score on FrontierMath (legacy) is 89.0% by GPT-5.6 Sol. 46 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	GPT-6 Astra	90.7%	estimated ± 2.3 pp, low confidence
2	GPT-5.6 Sol	89.0%	measured
3	GPT-5.6 Terra	84.9%	measured
4	GPT-5.6 Luna	78.6%	measured
5	GPT-5.5 Pro	52.4%	measured
6	GPT-5.5	51.7%	measured
7	GPT-5.4 Pro	50.0%	measured
8	GPT-5.4	48.1%	estimated ± 0.7 pp, low confidence
9	Claude Opus 4.8	47.7%	estimated ± 0.7 pp, low confidence
10	Claude Opus 4.7	44.3%	estimated ± 0.7 pp, low confidence
11	Claude Opus 4.7 (Adaptive)	43.8%	measured
12	Claude Opus 4.6	41.3%	estimated ± 0.7 pp, low confidence
13	GPT-5.2	41.3%	estimated ± 0.7 pp, low confidence
14	Muse Spark	39.6%	estimated ± 0.7 pp, low confidence
15	Gemini 3.5 Flash	39.6%	estimated ± 0.7 pp, low confidence
16	Kimi K2.6	39.6%	estimated ± 0.7 pp, low confidence
17	Gemini 3 Pro	38.2%	estimated ± 0.7 pp, low confidence
18	Gemini 3.1 Pro	37.5%	estimated ± 0.7 pp, low confidence
19	Gemini 3 Flash	36.3%	estimated ± 0.7 pp, low confidence
20	GLM-5.1	34.1%	estimated ± 0.7 pp, low confidence
21	Claude Sonnet 4.6	33.1%	estimated ± 0.7 pp, low confidence
22	GPT-5.1	31.8%	estimated ± 0.7 pp, low confidence
23	GPT-5.4 mini	29.0%	estimated ± 0.7 pp, low confidence
24	Kimi K2.5	28.7%	estimated ± 0.7 pp, low confidence
25	Qwen3.6 Plus	27.0%	estimated ± 0.7 pp, low confidence
26	GPT-5.4 nano	26.7%	estimated ± 0.7 pp, low confidence
27	o4-mini (high)	25.6%	estimated ± 0.7 pp, low confidence
28	Qwen 3.6 Max (preview)	23.9%	estimated ± 0.7 pp, low confidence
29	DeepSeek V3.2	22.9%	estimated ± 0.7 pp, low confidence
30	Kimi K2	22.3%	estimated ± 0.7 pp, low confidence
31	Qwen3.5 Plus	21.9%	estimated ± 0.7 pp, low confidence
32	Claude Opus 4.5	21.6%	estimated ± 0.7 pp, low confidence
33	Grok 4	20.5%	estimated ± 0.7 pp, low confidence
34	o3	19.6%	estimated ± 0.7 pp, low confidence
35	GLM-5	17.4%	estimated ± 0.7 pp, low confidence
36	Gemini 2.5 Pro	15.1%	estimated ± 0.7 pp, low confidence
37	Claude Sonnet 4.5	14.5%	estimated ± 0.7 pp, low confidence
38	o1	10.3%	estimated ± 0.7 pp, low confidence
39	Qwen3 235B 2507 (Reasoning)	9.5%	estimated ± 0.7 pp, low confidence
40	Qwen3.5 Flash	7.3%	estimated ± 0.7 pp, low confidence
41	Claude Haiku 4.5	7.0%	estimated ± 0.7 pp, low confidence
42	GPT-4.1	6.6%	estimated ± 0.7 pp, low confidence
43	Gemini 2.5 Flash	5.9%	estimated ± 0.7 pp, low confidence
44	GPT-4.1 mini	5.6%	estimated ± 0.7 pp, low confidence
45	GLM-4.6	4.9%	estimated ± 0.7 pp, low confidence
46	Grok 3 [Beta]	4.9%	estimated ± 0.7 pp, low confidence
47	GLM-4.7	3.6%	estimated ± 0.7 pp, low confidence
48	Claude 3.5 Sonnet	3.2%	estimated ± 0.7 pp, low confidence
49	DeepSeek V3	2.8%	estimated ± 0.7 pp, low confidence
50	GPT-4.1 nano	2.2%	estimated ± 0.7 pp, low confidence
51	Llama 4 Maverick	1.8%	estimated ± 0.7 pp, low confidence
52	GPT-4o	1.5%	estimated ± 0.7 pp, low confidence
53	Llama 4 Scout	1.1%	estimated ± 0.7 pp, low confidence

FrontierMath (legacy) leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General