Benchmark

MATH

Mathematics Aptitude Test of Heuristics

High school and competition math problems across algebra, geometry, number theory.

Models Tested

DeepSeek R1

Top Model

97.3%

Top Score

100 %

Max Possible

#	Model	MATH Score	Provider	Composite
1	DeepSeek R1 DeepSeek	97.3%	DeepSeek	92.3
2	Gemini 2.5 Pro Google	97.2%	Google	97.5
3	o3 OpenAI	96.7%	OpenAI	97.1
4	Claude Opus 4 Anthropic	96.1%	Anthropic	97.8
5	o4-mini OpenAI	93.2%	OpenAI	91.5
6	DeepSeek V3 DeepSeek	90.2%	DeepSeek	86.1
7	Qwen 3 235B Alibaba	88.5%	Alibaba	84.5
8	Grok 3 xAI	88.1%	xAI	88.8
9	Gemini 2.5 Flash Google	87.4%	Google	86.7
10	Claude Sonnet 4 Anthropic	85.6%	Anthropic	90.3
11	Llama 4 Maverick Meta	84.2%	Meta	87.5
12	Claude 3.5 Sonnet Anthropic	78.3%	Anthropic	88.4
13	GPT-4o OpenAI	76.6%	OpenAI	89.2
14	Llama 3.1 405B Meta	73.8%	Meta	82.1
15	Gemini 2.0 Flash Google	73.5%	Google	80.5
16	Llama 3.3 70B Meta	72.1%	Meta	79.8
17	Qwen 2.5 72B Alibaba	71.5%	Alibaba	76
18	Claude 3.5 Haiku Anthropic	69.7%	Anthropic	79.2
19	GPT-4o Mini OpenAI	69.5%	OpenAI	75.8
20	Mistral Large Mistral AI	69.2%	Mistral AI	76.5
21	Grok 2 xAI	68.5%	xAI	74.8
22	Llama 3.1 70B Meta	68%	Meta	74.3
23	Phi-4 Microsoft	62.5%	Microsoft	64.5
24	Mixtral 8x22B Mistral AI	60.4%	Mistral AI	65.2
25	Command R+ Cohere	55.3%	Cohere	60.8

Source: Hendrycks et al., 2021