Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-MilitaryAndWarfare

A measure of model refusal for Violence & Extremism (Level-1: Content Safety Risks, Level-2: Violence & Extremism) related to military and warfare. Includes Level-4 risks related to military operations and warfare.
Source:

Model Performance

#1
OpenAI-O1-1217
100.0%
#2
Claude-3-Sonnet
96.7%
#3
OpenAI-O4-mini-medium
95.0%
#4
Claude-3.5-Sonnet-1022
95.0%
#5
OpenAI-O3-medium
93.3%
#6
Kimi-K2-Instruct
91.7%
#7
Claude-3-Opus
90.0%
#8
GPT-OSS-120B
90.0%
#9
Gemini-2.5-Pro-0325
88.3%
#10
Gemini-2.5-Pro-Thinking
88.3%
#11
Claude-3.7-Sonnet
86.7%
#12
Claude-3.7-Sonnet-Thinking
86.7%
#13
Granite-3.0
86.7%
#14
OpenAI-O3-mini-medium
85.0%
#15
Gemini-2.5-Flash
85.0%
#16
Claude-4.0-Opus
83.3%
#17
Claude-4.0-Sonnet
83.3%
#18
Claude-4.0-Opus-Thinking
83.3%
#19
Llama-4-Maverick-17B
83.3%
#20
Llama-3.1-405B
83.3%
#21
Claude-4.0-Sonnet-Thinking
83.3%
#22
GPT-4.1
83.3%
#23
GPT-4.5
81.7%
#24
GPT-4-mini
80.0%
#25
Gemini-2.0-Pro-0121
80.0%
#26
Gemini-2.0-Flash
78.3%
#27
Gemini-1.5-Pro
73.3%
#28
Qwen-3
51.7%
#29
Qwen-3-Thinking
51.7%
#30
Grok-3-Beta
48.3%
#31
GPT-4o-0513
46.7%
#32
OpenAI-O1-mini
40.0%
#33
Grok-3-Mini-Beta
15.0%
#34
DeepSeek-R1
15.0%
#35
Grok-4-Thinking
13.3%
#36
Grok-4
13.3%
#37
DeepSeek-V3
10.0%
#38
Cohere-Command-R-Plus
10.0%
#39
GPT-3.5-Turbo
10.0%
#40
Mistral-Large-2
10.0%