Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-WeaponUsageDevelopment

A measure of model refusal for Violence & Extremism (Level-1: Content Safety Risks, Level-2: Violence & Extremism) related to weapon usage and development. Includes Level-4 risks like guns, explosives/dangerous materials, bioweapons/viruses/gain-of-function, nuclear weapons, chemical weapons, and radiological weapons.
Source:

Model Performance

#1
OpenAI-O3-medium
100.0%
#2
GPT-OSS-120B
98.9%
#3
Claude-3.5-Sonnet-1022
97.2%
#4
Claude-3-Opus
96.7%
#5
Claude-3-Sonnet
96.7%
#6
OpenAI-O1-1217
96.7%
#7
OpenAI-O4-mini-medium
95.6%
#8
OpenAI-O3-mini-medium
90.0%
#9
Kimi-K2-Instruct
84.4%
#10
Granite-3.0
83.3%
#11
Gemini-2.5-Pro-0325
83.3%
#12
Gemini-2.5-Pro-Thinking
83.3%
#13
GPT-4.5
82.2%
#14
Llama-4-Maverick-17B
82.2%
#15
GPT-4-mini
81.1%
#16
Gemini-1.5-Pro
78.9%
#17
Claude-4.0-Opus
78.3%
#18
Claude-4.0-Sonnet-Thinking
78.3%
#19
Claude-4.0-Opus-Thinking
78.3%
#20
Claude-4.0-Sonnet
78.3%
#21
Gemini-2.0-Flash
76.7%
#22
GPT-4.1
75.6%
#23
Claude-3.7-Sonnet-Thinking
75.0%
#24
Claude-3.7-Sonnet
75.0%
#25
Gemini-2.0-Pro-0121
70.6%
#26
Gemini-2.5-Flash
70.0%
#27
Llama-3.1-405B
62.2%
#28
GPT-4o-0513
60.0%
#29
Grok-3-Beta
59.4%
#30
Grok-4
55.6%
#31
Grok-4-Thinking
55.6%
#32
Qwen-3
48.9%
#33
Qwen-3-Thinking
48.9%
#34
Grok-3-Mini-Beta
47.8%
#35
GPT-3.5-Turbo
47.2%
#36
OpenAI-O1-mini
36.7%
#37
DeepSeek-R1
34.4%
#38
DeepSeek-V3
32.2%
#39
Mistral-Large-2
31.1%
#40
Cohere-Command-R-Plus
11.1%