Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-OffensiveLanguage

A measure of model refusal for Hate/Toxicity (Level-1: Content Safety Risks, Level-2: Hate/Toxicity) related to offensive language. Includes Level-4 risks like vulgarity and derogatory comments.
Source:

Model Performance

#1
Claude-3.5-Sonnet-1022
97.6%
#2
Claude-3.7-Sonnet
90.5%
#3
Claude-3.7-Sonnet-Thinking
90.5%
#4
Granite-3.0
90.5%
#5
Claude-4.0-Sonnet-Thinking
88.1%
#6
Claude-4.0-Sonnet
88.1%
#7
Claude-3-Sonnet
86.9%
#8
Claude-4.0-Opus
83.3%
#9
Claude-4.0-Opus-Thinking
83.3%
#10
Claude-3-Opus
81.0%
#11
GPT-OSS-120B
79.8%
#12
Gemini-2.0-Pro-0121
76.2%
#13
Llama-4-Maverick-17B
73.8%
#14
Gemini-2.5-Pro-0325
69.0%
#15
Gemini-2.5-Pro-Thinking
69.0%
#16
GPT-3.5-Turbo
67.9%
#17
Gemini-2.0-Flash
66.7%
#18
GPT-4.5
65.5%
#19
DeepSeek-R1
63.1%
#20
Gemini-2.5-Flash
63.1%
#21
GPT-4o-0513
59.5%
#22
Gemini-1.5-Pro
58.3%
#23
Qwen-3
57.1%
#24
Kimi-K2-Instruct
57.1%
#25
Qwen-3-Thinking
57.1%
#26
OpenAI-O3-mini-medium
52.4%
#27
GPT-4.1
50.0%
#28
Grok-3-Beta
48.8%
#29
OpenAI-O1-mini
47.6%
#30
OpenAI-O3-medium
45.2%
#31
DeepSeek-V3
45.2%
#32
GPT-4-mini
45.2%
#33
Grok-3-Mini-Beta
40.5%
#34
Llama-3.1-405B
39.3%
#35
OpenAI-O4-mini-medium
39.3%
#36
OpenAI-O1-1217
36.9%
#37
Mistral-Large-2
29.8%
#38
Cohere-Command-R-Plus
27.4%
#39
Grok-4-Thinking
22.6%
#40
Grok-4
22.6%