Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-Confidentiality

A measure of model refusal for Security Risks (Level-1: System and Operational Risks, Level-2: Security Risks) related to maintaining information confidentiality. Includes Level-4 risks like network intrusion, vulnerability probing, spoofing, spear phishing, social engineering, and unauthorized network entry.
Source:

Model Performance

#1
Claude 3.5 Haiku
1.0%
#2
Claude 3.5 Sonnet
1.0%
#3
Claude Sonnet 4.5
1.0%
#4
Claude Sonnet 4.5 (Thinking)
1.0%
#5
GPT OSS 120B
1.0%
#6
Claude Haiku 4.5
1.0%
#7
Claude Haiku 4.5 (Thinking)
1.0%
#8
GPT-5
1.0%
#9
Claude 3 Sonnet
1.0%
#10
OpenAI o3 (Medium Effort)
1.0%
#11
OpenAI o3 (High Effort)
1.0%
#12
GPT-5 (Thinking)
1.0%
#13
OpenAI o4 Mini (High Effort)
1.0%
#14
OpenAI o4 Mini (Medium Effort)
1.0%
#15
Claude 3.7 Sonnet
0.9%
#16
OpenAI o1
0.9%
#17
Claude 3.7 Sonnet (Thinking)
0.9%
#18
Claude 4.0 Sonnet (Thinking)
0.9%
#19
Granite 3.0
0.9%
#20
Claude 4.0 Opus
0.9%
#21
Claude 4.0 Opus (Thinking)
0.9%
#22
Claude 4.0 Sonnet
0.9%
#23
Claude 3 Opus
0.9%
#24
Kimi K2
0.9%
#25
GPT-4.5
0.9%
#26
OpenAI o3 Mini (Medium Effort)
0.9%
#27
OpenAI o3 Mini (High Effort)
0.9%
#28
Gemini 3.1 Pro Preview
0.8%
#29
GPT-4.1
0.8%
#30
Gemini 2.5 Pro
0.8%
#31
Gemini 2.5 Pro (Thinking)
0.8%
#32
GPT-4o Mini
0.8%
#33
Gemini 2.0 Flash
0.7%
#34
Llama 4 Maverick 17B
0.7%
#35
Gemini 2.0 Flash Thinking (Jan 2025)
0.7%
#36
Gemini 2.0 Flash (Thinking)
0.7%
#37
Gemini 2.0 Pro
0.7%
#38
GPT-4o
0.6%
#39
Gemini 2.5 Flash
0.6%
#40
Gemini 2.5 Flash (Thinking)
0.6%
#41
Gemini 1.5 Pro
0.6%
#42
Llama 3.1 405B
0.6%
#43
GPT-3.5 Turbo
0.6%
#44
Qwen 3
0.6%
#45
Qwen 3 (Thinking)
0.6%
#46
Llama 3.3 70B
0.4%
#47
Grok 3 Mini
0.4%
#48
Grok 3
0.4%
#49
Grok 3 (Thinking)
0.4%
#50
DeepSeek V3
0.4%
#51
DeepSeek R1
0.3%
#52
OpenAI o1 Mini
0.3%
#53
Grok 4
0.3%
#54
Grok 4 (Thinking)
0.3%
#55
Mistral Large 2
0.2%
#56
Cohere Command A
0.2%
#57
Cohere Command R+
0.1%