Benchmark

MMLU

Massive Multitask Language Understanding

57-subject test of world knowledge and problem solving across STEM, humanities, social sciences.

Models Tested

Claude Opus 4

Top Model

92.1%

Top Score

100 %

Max Possible

#	Model	MMLU Score	Provider	Composite
1	Claude Opus 4 Anthropic	92.1%	Anthropic	97.8
2	Gemini 2.5 Pro Google	92%	Google	97.5
3	o3 OpenAI	91.2%	OpenAI	97.1
4	DeepSeek R1 DeepSeek	90.8%	DeepSeek	92.3
5	Claude Sonnet 4 Anthropic	90.4%	Anthropic	90.3
6	Llama 4 Maverick Meta	89.5%	Meta	87.5
7	Grok 3 xAI	89.2%	xAI	88.8
8	GPT-4o OpenAI	88.7%	OpenAI	89.2
9	Claude 3.5 Sonnet Anthropic	88.7%	Anthropic	88.4
10	DeepSeek V3 DeepSeek	88.5%	DeepSeek	86.1
11	o4-mini OpenAI	88.3%	OpenAI	91.5
12	Qwen 3 235B Alibaba	87.8%	Alibaba	84.5
13	Llama 3.1 405B Meta	87.3%	Meta	82.1
14	Gemini 2.5 Flash Google	86.5%	Google	86.7
15	Llama 3.3 70B Meta	86%	Meta	79.8
16	Qwen 2.5 72B Alibaba	85.2%	Alibaba	76
17	Gemini 2.0 Flash Google	85.1%	Google	80.5
18	Claude 3.5 Haiku Anthropic	84.1%	Anthropic	79.2
19	Mistral Large Mistral AI	84%	Mistral AI	76.5
20	Llama 3.1 70B Meta	83.7%	Meta	74.3
21	Grok 2 xAI	83.2%	xAI	74.8
22	GPT-4o Mini OpenAI	82%	OpenAI	75.8
23	Phi-4 Microsoft	78.5%	Microsoft	64.5
24	Mixtral 8x22B Mistral AI	77.8%	Mistral AI	65.2
25	Command R+ Cohere	75.6%	Cohere	60.8

Source: Hendrycks et al., 2021