Multilingual

MMLU-ProX leaderboard

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

As of 2026-10-07, the highest measured score on MMLU-ProX is 87.0% by Qwen3.7 Max. 38 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	Claude Opus 5.5	91.7%	estimated ± 1.7 pp, low confidence
2	Claude Mythos 5	91.1%	estimated ± 1.7 pp, low confidence
3	Claude Sonnet 5.5	90.4%	estimated ± 1.7 pp, low confidence
4	Claude Opus 5	90.1%	estimated ± 1.7 pp, low confidence
5	Claude Fable 5.1	90.0%	estimated ± 1.7 pp, low confidence
6	Claude Opus 4.8	88.3%	estimated ± 1.7 pp, low confidence
7	Hy4 preview	87.7%	estimated ± 1.7 pp, low confidence
8	Qwen3.8-Flash-Next	87.0%	estimated ± 1.7 pp, low confidence
9	Qwen3.7 Max	87.0%	measured
10	Qwen3.8-Omni-Flash	86.8%	estimated ± 1.7 pp, low confidence
11	Composer 2.5	86.6%	estimated ± 1.7 pp, low confidence
12	Ornith-1.5-397B	86.5%	estimated ± 1.7 pp, low confidence
13	Ornith-1.0-397B	86.3%	estimated ± 1.7 pp, low confidence
14	Laguna S 2.1	86.1%	estimated ± 1.7 pp, low confidence
15	Claude Sonnet 5	86.0%	estimated ± 1.7 pp, medium confidence
16	Grok 4.5	85.9%	estimated ± 1.7 pp, medium confidence
17	Beam	85.9%	estimated ± 1.7 pp, medium confidence
18	SWE-1.7	85.8%	estimated ± 1.7 pp, medium confidence
19	Claude Opus 4.5	85.7%	measured
20	Kimi K2.6	85.4%	estimated ± 1.7 pp, medium confidence
21	Qwen3.7 Plus	85.4%	measured
22	MiniMax M2.7	85.3%	estimated ± 1.7 pp, medium confidence
23	DeepSeek V4 Pro 0813	85.2%	estimated ± 1.7 pp, medium confidence
24	dots3-note Preview	85.0%	estimated ± 1.7 pp, medium confidence
25	Qwen3.5 397B	84.7%	measured
26	Qwen3.6 Plus	84.7%	measured
27	Composer 2	84.3%	estimated ± 1.7 pp, medium confidence
28	DeepSeek V4 Flash 0731	84.1%	estimated ± 1.7 pp, medium confidence
29	Ling 3.0 Flash	83.8%	estimated ± 1.7 pp, medium confidence
30	Ornith-1.5-35B-A3B	83.4%	estimated ± 1.7 pp, medium confidence
31	Qwen3.6-27B	83.3%	estimated ± 1.7 pp, medium confidence
32	GLM-5	83.1%	measured
33	Nemotron 3 Ultra	83.0%	measured
34	Ornith-1.0-35B	82.6%	estimated ± 1.7 pp, medium confidence
35	Kimi K2.5	82.3%	measured
36	Qwen3.5-122B-A10B	82.2%	measured
37	Qwen3.5-27B	82.2%	measured
38	Qwen3.6-35B-A3B	81.7%	estimated ± 1.7 pp, low confidence
39	Qwen3.5-35B-A3B	81.0%	measured
40	Laguna M.1	80.1%	estimated ± 1.7 pp, low confidence
41	Laguna XS 2.1	80.1%	estimated ± 1.7 pp, low confidence
42	Qwen3 235B 2507	79.4%	measured
43	LongCat-Flash-Lite-Sparse	78.6%	estimated ± 1.7 pp, low confidence
44	Laguna XS.2	77.9%	estimated ± 1.7 pp, low confidence
45	Ornith-1.5-9B	76.5%	estimated ± 1.7 pp, low confidence
46	Ornith-1.0-9B	75.5%	estimated ± 1.7 pp, low confidence
47	GPT-4.1	72.7%	measured
48	Granite 4.2 30B	71.1%	estimated ± 1.7 pp, low confidence
49	DeepSeek V3 0324	70.5%	measured
50	Nemotron 3.5 Lightning 30B A3B NVFP4	68.6%	estimated ± 1.7 pp, low confidence
51	Granite 4.2 8B	65.9%	estimated ± 1.7 pp, low confidence
52	LLaDA2.2-flash	63.2%	estimated ± 1.7 pp, low confidence
53	GPT-4o	61.1%	measured
54	Phi-4	49.9%	measured

MMLU-ProX leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General