Benchmark

IFEval

Instruction Following Evaluation

Strict instruction-following compliance across formatting and constraint tasks.

Models Tested

Claude Opus 4

Top Model

94.2%

Top Score

100 %

Max Possible

#	Model	IFEval Score	Provider	Composite
1	Claude Opus 4 Anthropic	94.2%	Anthropic	97.8
2	o3 OpenAI	92.8%	OpenAI	97.1
3	Gemini 2.5 Pro Google	92.3%	Google	97.5
4	Claude Sonnet 4 Anthropic	90.1%	Anthropic	90.3
5	o4-mini OpenAI	89.5%	OpenAI	91.5
6	DeepSeek R1 DeepSeek	88.3%	DeepSeek	92.3
7	Grok 3 xAI	87.8%	xAI	88.8
8	Llama 4 Maverick Meta	87.4%	Meta	87.5
9	DeepSeek V3 DeepSeek	86.7%	DeepSeek	86.1
10	Claude 3.5 Sonnet Anthropic	86.5%	Anthropic	88.4
11	GPT-4o OpenAI	86.1%	OpenAI	89.2
12	Gemini 2.5 Flash Google	85.7%	Google	86.7
13	Qwen 3 235B Alibaba	85.2%	Alibaba	84.5
14	Llama 3.1 405B Meta	84.3%	Meta	82.1
15	Llama 3.3 70B Meta	83.5%	Meta	79.8
16	Gemini 2.0 Flash Google	83.2%	Google	80.5
17	GPT-4o Mini OpenAI	82.4%	OpenAI	75.8
18	Claude 3.5 Haiku Anthropic	82.1%	Anthropic	79.2
19	Mistral Large Mistral AI	81.2%	Mistral AI	76.5
20	Qwen 2.5 72B Alibaba	80.1%	Alibaba	76
21	Llama 3.1 70B Meta	79.8%	Meta	74.3
22	Grok 2 xAI	78.5%	xAI	74.8
23	Mixtral 8x22B Mistral AI	75.3%	Mistral AI	65.2
24	Command R+ Cohere	72.5%	Cohere	60.8
25	Phi-4 Microsoft	70.2%	Microsoft	64.5

Source: Zhou et al., 2023