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.

Model Performance

#1
Claude Haiku 4.5
100.0%
#2
Claude Haiku 4.5 (Thinking)
100.0%
#3
GPT-4.5
100.0%
#4
Grok 3 Mini
100.0%
#5
Grok 3
100.0%
#6
Gemini 3.1 Pro Preview
100.0%
#7
Grok 3 (Thinking)
100.0%
#8
GPT OSS 120B
100.0%
#9
Claude 3.5 Sonnet
98.9%
#10
Claude Sonnet 4.5
97.8%
#11
Claude Sonnet 4.5 (Thinking)
97.8%
#12
Claude 3 Sonnet
97.8%
#13
Claude 3 Opus
95.6%
#14
OpenAI o3 Mini (High Effort)
88.9%
#15
OpenAI o3 Mini (Medium Effort)
88.9%
#16
Claude 3.7 Sonnet (Thinking)
87.8%
#17
Claude 3.7 Sonnet
87.8%
#18
Kimi K2
83.3%
#19
Llama 4 Maverick 17B
82.2%
#20
GPT-4.1
77.8%
#21
Gemini 2.5 Flash (Thinking)
74.4%
#22
Gemini 2.5 Flash
74.4%
#23
Llama 3.1 405B
68.9%
#24
GPT-4o Mini
67.8%
#25
Gemini 2.0 Flash (Thinking)
63.3%
#26
Gemini 2.0 Flash
63.3%
#27
Gemini 2.0 Flash Thinking (Jan 2025)
63.3%
#28
OpenAI o4 Mini (High Effort)
60.0%
#29
Gemini 2.0 Pro
60.0%
#30
OpenAI o4 Mini (Medium Effort)
60.0%
#31
Granite 3.0
57.6%
#32
Qwen 3 (Thinking)
56.7%
#33
Qwen 3
56.7%
#34
GPT-3.5 Turbo
54.8%
#35
GPT-4o
48.9%
#36
Gemini 1.5 Pro
44.4%
#37
OpenAI o1 Mini
37.8%
#38
DeepSeek R1
35.6%
#39
Grok 4
33.3%
#40
Grok 4 (Thinking)
33.3%
#41
DeepSeek V3
28.9%
#42
Llama 3.3 70B
27.8%
#43
OpenAI o1
22.0%
#44
Cohere Command A
17.8%
#45
Claude 3.5 Haiku
16.7%
#46
Gemini 2.5 Pro (Thinking)
13.3%
#47
Claude 4.0 Sonnet
13.3%
#48
Claude 4.0 Sonnet (Thinking)
13.3%
#49
Gemini 2.5 Pro
13.3%
#50
OpenAI o3 (Medium Effort)
10.7%
#51
OpenAI o3 (High Effort)
10.7%
#52
GPT-5
8.0%
#53
GPT-5 (Thinking)
8.0%
#54
Mistral Large 2
6.7%
#55
Claude 4.0 Opus
6.0%
#56
Claude 4.0 Opus (Thinking)
6.0%
#57
Cohere Command R+
5.6%