Benchmark

LiveCodeBench

Live Code Benchmark

Contamination-free code generation from recent competitive programming problems.

Models Tested

Claude Opus 4

Top Model

73.8%

Top Score

100 %

Max Possible

#	Model	LiveCodeBench Score	Provider	Composite
1	Claude Opus 4 Anthropic	73.8%	Anthropic	97.8
2	o3 OpenAI	72.1%	OpenAI	97.1
3	Gemini 2.5 Pro Google	70.4%	Google	97.5
4	DeepSeek R1 DeepSeek	65.8%	DeepSeek	92.3
5	o4-mini OpenAI	65.3%	OpenAI	91.5
6	Claude Sonnet 4 Anthropic	55.1%	Anthropic	90.3
7	Grok 3 xAI	52.8%	xAI	88.8
8	Llama 4 Maverick Meta	50.2%	Meta	87.5
9	Gemini 2.5 Flash Google	48.6%	Google	86.7
10	DeepSeek V3 DeepSeek	43.5%	DeepSeek	86.1
11	Qwen 3 235B Alibaba	42.8%	Alibaba	84.5
12	Claude 3.5 Sonnet Anthropic	40.2%	Anthropic	88.4
13	GPT-4o OpenAI	33.1%	OpenAI	89.2
14	Claude 3.5 Haiku Anthropic	29.3%	Anthropic	79.2
15	Gemini 2.0 Flash Google	28.5%	Google	80.5
16	Llama 3.1 405B Meta	27.6%	Meta	82.1
17	Llama 3.3 70B Meta	25.1%	Meta	79.8
18	Mistral Large Mistral AI	22.4%	Mistral AI	76.5
19	Qwen 2.5 72B Alibaba	21.3%	Alibaba	76
20	Grok 2 xAI	20.5%	xAI	74.8
21	Llama 3.1 70B Meta	19.8%	Meta	74.3
22	GPT-4o Mini OpenAI	18.3%	OpenAI	75.8
23	Mixtral 8x22B Mistral AI	12.1%	Mistral AI	65.2
24	Command R+ Cohere	9.2%	Cohere	60.8
25	Phi-4 Microsoft	8.5%	Microsoft	64.5

Source: Jain et al., 2024