⌨️ Code LLM 리더보드

Code generation, debugging, and software engineering tasks.

#	모델	종합	HumanEval	SWE-Bench	LiveCodeBench	속도
1	Claude Opus 4 Anthropic	97.8	94.5%	72.5%	73.8%	45 t/s
2	Gemini 2.5 Pro Google	97.5	93%	63.8%	70.4%	85 t/s
3	o3 OpenAI	97.1	92.7%	71.7%	72.1%	58 t/s
4	DeepSeek R1 DeepSeek	92.3	92.1%	49.2%	65.8%	32 t/s
5	o4-mini OpenAI	91.5	91.5%	68.1%	65.3%	145 t/s
6	Claude Sonnet 4 Anthropic	90.3	93.7%	62.3%	55.1%	79 t/s
7	GPT-4o OpenAI	89.2	90.2%	38.4%	33.1%	110 t/s
8	Grok 3 xAI	88.8	90.5%	55.2%	52.8%	55 t/s
9	Claude 3.5 Sonnet Anthropic	88.4	92%	49%	40.2%	82 t/s
10	Llama 4 Maverick Meta	87.5	90.8%	47.5%	50.2%	120 t/s
11	Gemini 2.5 Flash Google	86.7	89.8%	48.2%	48.6%	203 t/s
12	DeepSeek V3 DeepSeek	86.1	89.4%	42%	43.5%	68 t/s
13	Qwen 3 235B Alibaba	84.5	88.7%	41.3%	42.8%	95 t/s
14	Llama 3.1 405B Meta	82.1	89%	35.2%	27.6%	29 t/s
15	Gemini 2.0 Flash Google	80.5	87.6%	36.8%	28.5%	258 t/s
16	Llama 3.3 70B Meta	79.8	86.4%	32.8%	25.1%	65 t/s
17	Claude 3.5 Haiku Anthropic	79.2	88.1%	40.6%	29.3%	148 t/s
18	Mistral Large Mistral AI	76.5	84.3%	30.1%	22.4%	48 t/s
19	Qwen 2.5 72B Alibaba	76	82.4%	26.5%	21.3%	55 t/s
20	GPT-4o Mini OpenAI	75.8	87.2%	23.4%	18.3%	167 t/s
21	Grok 2 xAI	74.8	82.1%	28.3%	20.5%	60 t/s
22	Llama 3.1 70B Meta	74.3	80.5%	27.3%	19.8%	62 t/s
23	Mixtral 8x22B Mistral AI	65.2	75.8%	18.2%	12.1%	82 t/s
24	Phi-4 Microsoft	64.5	72.4%	12.5%	8.5%	125 t/s