Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-ViolentActs

A measure of model refusal for Violence & Extremism (Level-1: Content Safety Risks, Level-2: Violence & Extremism) related to violent acts. Includes Level-4 risks like violence against persons (including murder), animals, property damage, and environmental harm.
Source:

Model Performance

#1
Claude 3 Sonnet
100.0%
#2
Claude 3.5 Sonnet
100.0%
#3
Llama 4 Maverick 17B
100.0%
#4
Gemini 2.5 Flash (Thinking)
100.0%
#5
Gemini 2.5 Flash
100.0%
#6
OpenAI o1 Mini
100.0%
#7
Claude Haiku 4.5
98.9%
#8
Claude Haiku 4.5 (Thinking)
98.9%
#9
Claude 3 Opus
98.3%
#10
Granite 3.0
97.8%
#11
GPT OSS 120B
97.8%
#12
Claude Sonnet 4.5 (Thinking)
97.5%
#13
Claude Sonnet 4.5
97.5%
#14
OpenAI o4 Mini (High Effort)
96.7%
#15
OpenAI o4 Mini (Medium Effort)
96.7%
#16
Claude 3.7 Sonnet
95.8%
#17
Claude 3.7 Sonnet (Thinking)
95.8%
#18
OpenAI o3 Mini (High Effort)
95.0%
#19
OpenAI o3 Mini (Medium Effort)
95.0%
#20
Kimi K2
91.7%
#21
GPT-4o Mini
86.7%
#22
Gemini 2.0 Flash Thinking (Jan 2025)
86.7%
#23
Gemini 2.0 Flash
86.7%
#24
Gemini 2.0 Flash (Thinking)
86.7%
#25
Claude 4.0 Sonnet (Thinking)
84.8%
#26
Claude 4.0 Sonnet
84.8%
#27
GPT-4.1
83.3%
#28
Gemini 1.5 Pro
80.0%
#29
Claude 4.0 Opus
79.0%
#30
Claude 4.0 Opus (Thinking)
79.0%
#31
GPT-5 (Thinking)
73.9%
#32
GPT-5
73.9%
#33
Llama 3.1 405B
68.3%
#34
Qwen 3
66.7%
#35
Qwen 3 (Thinking)
66.7%
#36
Gemini 2.0 Pro
65.6%
#37
GPT-4o
65.0%
#38
Claude 3.5 Haiku
60.9%
#39
OpenAI o3 (High Effort)
60.0%
#40
OpenAI o3 (Medium Effort)
60.0%
#41
Gemini 2.5 Pro
59.4%
#42
Gemini 2.5 Pro (Thinking)
59.4%
#43
OpenAI o1
52.9%
#44
DeepSeek R1
52.5%
#45
GPT-4.5
50.0%
#46
DeepSeek V3
45.0%
#47
Gemini 3.1 Pro Preview
44.2%
#48
Llama 3.3 70B
41.7%
#49
Grok 3 Mini
38.9%
#50
GPT-3.5 Turbo
37.8%
#51
Cohere Command A
28.3%
#52
Grok 3 (Thinking)
27.8%
#53
Grok 3
27.8%
#54
Mistral Large 2
26.7%
#55
Cohere Command R+
25.0%
#56
Grok 4
13.3%
#57
Grok 4 (Thinking)
13.3%