Vision & documents

AA-MMMU-Pro leaderboard

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

As of 2026-10-07, the highest measured score on AA-MMMU-Pro is 87.7% by Claude Opus 5.5. 25 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	87.7%	measured
2	GPT-6 Astra	86.9%	measured
3	GPT-6.1 Sol	86.0%	measured
4	Gemini 3.8 Flash	85.6%	measured
5	Gemini 3.7 Flash	85.5%	measured
6	Claude Opus 5	84.7%	measured
7	Gemini 3.5 Flash	84.3%	measured
8	Claude Opus 4.8	83.7%	estimated ± 2.9 pp, high confidence
9	Hy4 preview	83.7%	estimated ± 2.9 pp, high confidence
10	Claude Sonnet 5.5	83.4%	estimated ± 2.9 pp, high confidence
11	GPT-5.6 Sol	83.4%	measured
12	Gemini 3.6 Flash	83.2%	measured
13	Claude Mythos 5	82.9%	estimated ± 3.2 pp, medium confidence
14	GPT-6 Sol	82.9%	measured
15	Qwen3.8 Max Preview	82.8%	measured
16	Gemini 3.1 Pro	82.4%	measured
17	GLM-5.3-Flash	81.9%	estimated ± 2.9 pp, high confidence
18	Seed 2.1 Pro	81.2%	estimated ± 2.2 pp, high confidence
19	Seed 2.1 Turbo	80.9%	estimated ± 2.2 pp, high confidence
20	GPT-5.6 Terra	80.7%	measured
21	Kimi K3	80.5%	measured
22	Muse Spark	80.5%	measured
23	Qwen3.7 Plus	80.5%	measured
24	Grok 4.5	80.4%	measured
25	Gemini 3 Pro	80.2%	measured
26	Claude Fable 5	80.2%	estimated ± 2.9 pp, high confidence
27	Qwen3.8 Max	79.9%	estimated ± 1.9 pp, medium confidence
28	GPT-5.5	79.9%	measured
29	Qwen3.8-Flash-Next	79.8%	measured
30	GPT-6 Luna	79.7%	measured
31	dots3-note Preview	79.7%	estimated ± 3.0 pp, medium confidence
32	Holo2-235B-A22B	79.6%	estimated ± 9.4 pp, medium confidence
33	Qwen3.8-Omni-Flash	79.5%	estimated ± 1.9 pp, medium confidence
34	Kimi K2.6	79.4%	measured
35	Apodex 1.1	79.2%	measured
36	Apodex 1.1 Mini	79.2%	measured
37	Gemini 3.5 Flash-Lite	79.0%	measured
38	Ling 3.0 Flash VL	79.0%	measured
39	Claude Opus 4.7 (Adaptive)	78.8%	measured
40	Gemini 3 Flash	78.6%	measured
41	GPT-5.6 Luna	78.6%	measured
42	MiniMax M3	78.6%	measured
43	GPT-5.3 Codex	78.5%	measured
44	GPT-5.4	78.4%	measured
45	Grok 4.3	78.1%	measured
46	Qwen3.6 Plus	78.0%	measured
47	GPT-5.2	77.6%	estimated ± 4.9 pp, medium confidence
48	Claude Sonnet 5	77.3%	measured
49	DeepSeek V4.1 Flash	77.0%	measured
50	Claude Opus 4.7	76.4%	measured
51	Mistral Large 4	76.4%	measured
52	Step 5 Preview	76.4%	measured
53	GPT-5.2-Codex	76.3%	measured
54	Qwen3.8-27B	76.3%	measured
55	Pareto 26.9	76.0%	estimated ± 4.9 pp, medium confidence
56	MiMo-V2.5	75.9%	estimated ± 4.9 pp, medium confidence
57	GPT-5.1	75.5%	measured
58	Claude Opus 4.6 (Adaptive)	75.4%	measured
59	Kimi K2.5	75.4%	measured
60	Kimi K2.5 (Reasoning)	75.4%	measured
61	Step 3.7 Flash	75.3%	measured
62	Qwen3.5-122B-A10B	75.0%	measured
63	Qwen3.5-27B	75.0%	measured
64	Qwen3.6-35B-A3B	75.0%	measured
65	Gemini 2.5 Pro	74.9%	measured
66	Qwen3.6-27B	74.6%	measured
67	GPT-5 (medium)	74.3%	measured
68	Muse Glimmer 30B	74.3%	measured
69	GPT-5 (high)	74.2%	measured
70	Claude Opus 4.5 Thinking	74.0%	measured
71	Inkling-Small	74.0%	measured
72	Inkling	73.5%	measured
73	Gemma 4 31B	73.4%	measured
74	GPT-5.4 mini	73.3%	measured
75	Grok 4.20	73.2%	estimated ± 2.2 pp, high confidence
76	MiMo-V2.6-Flash	73.1%	measured
77	GLM-5V-Turbo	72.8%	measured
78	Qwen3.5-35B-A3B	72.7%	measured
79	Claude Opus 4.6	72.5%	measured
80	GPT-5.1-Codex	72.5%	measured
81	GPT-5.1-Codex-Max	72.5%	measured
82	Claude Opus 4.5	71.2%	measured
83	Claude Sonnet 4.6	70.6%	measured
84	o3	70.1%	measured
85	MiMo-V2-Omni	69.9%	measured
86	Interfaze Beta	69.7%	estimated ± 4.9 pp, medium confidence
87	Gemma 4 12B	69.7%	measured
88	Gemma 4 26B A4B	69.2%	measured
89	Grok 4	68.8%	measured
90	Claude 4.1 Opus Thinking	67.9%	measured
91	Ternary Bonsai 2 27B	67.4%	estimated ± 1.9 pp, low confidence
92	Gemini 2.5 Flash	65.5%	measured
93	GPT-5.4 nano	65.4%	measured
94	Mistral Medium 3.5 128B	64.9%	measured
95	Grok 4.1 Fast (Reasoning)	63.3%	measured
96	Command A+	63.2%	measured
97	Claude 4 Sonnet	62.4%	measured
98	Llama 4 Maverick	62.1%	measured
99	Grok 4 Fast (Reasoning)	61.8%	measured
100	GPT-4.1	61.2%	measured
101	Holo2-30B-A3B	60.6%	estimated ± 9.4 pp, medium confidence
102	Qwen3-Omni-30B-A3B-Thinking	60.2%	measured
103	Holo2-4B	59.0%	estimated ± 9.4 pp, medium confidence
104	Holo2-8B	59.0%	estimated ± 9.4 pp, medium confidence
105	GPT-4.1 mini	58.7%	measured
106	Mistral Small 4	56.8%	measured
107	Mistral Small 4 (Reasoning)	56.8%	measured
108	Mistral Large 3	55.7%	measured
109	Qwen3-Omni-30B-A3B-Instruct	55.5%	measured
110	Gemini 1.5 Pro	55.0%	measured
111	Nemotron 3 Nano Omni 30B A3B	53.2%	measured
112	Mistral Medium 3	53.0%	measured
113	Llama 4 Scout	52.9%	measured
114	Qwen3.5 397B	52.7%	measured
115	Qwen3.5 397B (Reasoning)	52.7%	measured
116	Gemma 4 E4B	51.4%	measured
117	Grok 4.1 Fast	48.4%	measured
118	Gemma 3 27B	48.0%	measured
119	Gemma 4 E2B	44.6%	measured
120	Nova Pro	44.3%	measured
121	GPT-4o mini	41.5%	measured
122	GPT-4.1 nano	40.1%	measured
123	Claude 3 Haiku	30.8%	measured
124	North Micro Vision Instruct	29.3%	estimated ± 1.9 pp, low confidence
125	LFM2.5-VL-1.6B-Extract	26.5%	measured
126	LFM2.5-VL-3B	16.2%	estimated ± 1.2 pp, medium confidence
127	ZAYA1-VL-8B	15.2%	estimated ± 1.2 pp, medium confidence
128	Phi-4 Multimodal Instruct	14.5%	measured
129	LFM2.5-VL-450M	13.4%	estimated ± 1.2 pp, medium confidence

AA-MMMU-Pro leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General