Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-IllegalServicesExploitation

A measure of model refusal for Criminal Activities (Level-1: Legal and Rights-Related Risks, Level-2: Criminal Activities) related to illegal services/exploitation. Includes Level-4 risks like human trafficking, sexual exploitation, and prostitution.

Model Performance

#1
Claude 3 Opus
100.0%
#2
Claude 3.5 Haiku
100.0%
#3
Claude 3.5 Sonnet
100.0%
#4
Claude 3.7 Sonnet
100.0%
#5
Claude 3.7 Sonnet (Thinking)
100.0%
#6
Claude Sonnet 4.5 (Thinking)
100.0%
#7
Claude Sonnet 4.5
100.0%
#8
Claude Haiku 4.5 (Thinking)
100.0%
#9
Claude Haiku 4.5
100.0%
#10
GPT-4.1
100.0%
#11
OpenAI o3 (High Effort)
100.0%
#12
OpenAI o3 (Medium Effort)
100.0%
#13
GPT OSS 120B
100.0%
#14
Granite 3.0
97.3%
#15
Claude 4.0 Sonnet (Thinking)
96.7%
#16
Claude 4.0 Sonnet
96.7%
#17
OpenAI o1
96.0%
#18
OpenAI o3 Mini (High Effort)
94.7%
#19
OpenAI o3 Mini (Medium Effort)
94.7%
#20
OpenAI o4 Mini (High Effort)
93.3%
#21
OpenAI o4 Mini (Medium Effort)
93.3%
#22
GPT-5
90.2%
#23
GPT-5 (Thinking)
90.2%
#24
Gemini 2.5 Flash
89.3%
#25
Gemini 2.5 Flash (Thinking)
89.3%
#26
Kimi K2
86.7%
#27
Claude 4.0 Opus
85.5%
#28
Claude 4.0 Opus (Thinking)
85.5%
#29
Gemini 2.0 Flash Thinking (Jan 2025)
84.4%
#30
Gemini 2.0 Flash
84.4%
#31
Gemini 2.0 Flash (Thinking)
84.4%
#32
Gemini 3.1 Pro Preview
80.0%
#33
Claude 3 Sonnet
73.9%
#34
GPT-4o Mini
73.3%
#35
Gemini 2.5 Pro
73.0%
#36
Gemini 2.5 Pro (Thinking)
73.0%
#37
Gemini 1.5 Pro
72.8%
#38
GPT-3.5 Turbo
69.3%
#39
Llama 3.1 405B
66.7%
#40
GPT-4o
62.2%
#41
Grok 3 Mini
60.2%
#42
GPT-4.5
60.0%
#43
Qwen 3
60.0%
#44
Qwen 3 (Thinking)
60.0%
#45
Llama 4 Maverick 17B
58.9%
#46
Grok 4 (Thinking)
55.6%
#47
Grok 4
55.6%
#48
Gemini 2.0 Pro
54.6%
#49
Grok 3
50.0%
#50
Grok 3 (Thinking)
50.0%
#51
Llama 3.3 70B
48.9%
#52
DeepSeek R1
41.1%
#53
DeepSeek V3
37.8%
#54
OpenAI o1 Mini
27.8%
#55
Mistral Large 2
22.2%
#56
Cohere Command A
20.0%
#57
Cohere Command R+
15.6%