Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-SupportingMaliciousOperations

A measure of model refusal for Violence & Extremism (Level-1: Content Safety Risks, Level-2: Violence & Extremism) related to supporting malicious organized groups. Includes Level-4 risks like extremism, terrorism, and criminal organization support.
Source:

Model Performance

#1
Claude-3.5-Sonnet-1022
98.9%
#2
Claude-3-Sonnet
97.8%
#3
Granite-3.0
97.8%
#4
GPT-OSS-120B
97.8%
#5
OpenAI-O3-medium
97.8%
#6
OpenAI-O4-mini-medium
96.7%
#7
Claude-3-Opus
95.6%
#8
Claude-4.0-Opus
95.6%
#9
Claude-4.0-Sonnet-Thinking
95.6%
#10
Claude-4.0-Sonnet
95.6%
#11
Claude-4.0-Opus-Thinking
95.6%
#12
OpenAI-O1-1217
94.4%
#13
OpenAI-O3-mini-medium
88.9%
#14
Gemini-2.5-Pro-0325
88.9%
#15
Gemini-2.5-Pro-Thinking
88.9%
#16
Claude-3.7-Sonnet
87.8%
#17
Claude-3.7-Sonnet-Thinking
87.8%
#18
GPT-4.5
87.8%
#19
Kimi-K2-Instruct
83.3%
#20
Llama-4-Maverick-17B
82.2%
#21
GPT-4.1
77.8%
#22
Gemini-2.5-Flash
74.4%
#23
Llama-3.1-405B
68.9%
#24
GPT-4-mini
67.8%
#25
Gemini-2.0-Pro-0121
65.6%
#26
Gemini-2.0-Flash
63.3%
#27
Qwen-3
56.7%
#28
Qwen-3-Thinking
56.7%
#29
GPT-4o-0513
48.9%
#30
Gemini-1.5-Pro
44.4%
#31
Grok-3-Mini-Beta
38.9%
#32
GPT-3.5-Turbo
37.8%
#33
OpenAI-O1-mini
37.8%
#34
DeepSeek-R1
35.6%
#35
DeepSeek-V3
28.9%
#36
Grok-3-Beta
27.8%
#37
Grok-4-Thinking
13.3%
#38
Grok-4
13.3%
#39
Mistral-Large-2
6.7%
#40
Cohere-Command-R-Plus
5.6%