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-3.5-Sonnet-1022
100.0%
#2
GPT-OSS-120B
100.0%
#3
OpenAI-O3-medium
100.0%
#4
Claude-3-Opus
98.3%
#5
Claude-3-Sonnet
98.3%
#6
Granite-3.0
96.7%
#7
OpenAI-O1-1217
96.7%
#8
OpenAI-O4-mini-medium
96.7%
#9
Claude-3.7-Sonnet
95.8%
#10
Claude-3.7-Sonnet-Thinking
95.8%
#11
OpenAI-O3-mini-medium
95.0%
#12
Claude-4.0-Sonnet
94.2%
#13
Claude-4.0-Sonnet-Thinking
94.2%
#14
Claude-4.0-Opus
93.3%
#15
Claude-4.0-Opus-Thinking
93.3%
#16
GPT-4.5
91.7%
#17
Kimi-K2-Instruct
91.7%
#18
Gemini-2.5-Pro-0325
88.3%
#19
Gemini-2.0-Pro-0121
88.3%
#20
Gemini-2.5-Pro-Thinking
88.3%
#21
Gemini-2.5-Flash
86.7%
#22
Gemini-2.0-Flash
86.7%
#23
GPT-4-mini
86.7%
#24
Llama-4-Maverick-17B
85.0%
#25
GPT-4.1
83.3%
#26
Gemini-1.5-Pro
81.7%
#27
Llama-3.1-405B
68.3%
#28
Qwen-3
66.7%
#29
Qwen-3-Thinking
66.7%
#30
GPT-3.5-Turbo
65.0%
#31
GPT-4o-0513
65.0%
#32
Grok-3-Mini-Beta
60.0%
#33
DeepSeek-R1
52.5%
#34
Grok-4
51.7%
#35
Grok-4-Thinking
51.7%
#36
Grok-3-Beta
48.3%
#37
DeepSeek-V3
45.0%
#38
OpenAI-O1-mini
40.0%
#39
Mistral-Large-2
26.7%
#40
Cohere-Command-R-Plus
25.0%