Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-SuicidalAndNonsuicidalSelfinjury

A measure of model refusal for Self-harm (Level-1: Content Safety Risks, Level-2: Self-harm) related to suicidal and non-suicidal self-injury. Includes Level-4 risks like suicide, cutting, and eating disorders (anorexia/bulimia).
Source:

Model Performance

#1
Claude-3-Sonnet
100.0%
#2
Claude-3.5-Sonnet-1022
100.0%
#3
OpenAI-O1-1217
100.0%
#4
OpenAI-O4-mini-medium
100.0%
#5
OpenAI-O3-medium
100.0%
#6
GPT-OSS-120B
100.0%
#7
Gemini-2.5-Pro-0325
100.0%
#8
Gemini-2.5-Pro-Thinking
100.0%
#9
Claude-3.7-Sonnet
97.8%
#10
Claude-3-Opus
97.8%
#11
Claude-4.0-Sonnet
97.8%
#12
Claude-3.7-Sonnet-Thinking
97.8%
#13
GPT-4.5
97.8%
#14
Claude-4.0-Sonnet-Thinking
97.8%
#15
Claude-4.0-Opus-Thinking
96.7%
#16
Claude-4.0-Opus
96.7%
#17
Gemini-2.5-Flash
95.6%
#18
Gemini-2.0-Flash
95.6%
#19
Gemini-2.0-Pro-0121
93.3%
#20
OpenAI-O3-mini-medium
93.3%
#21
Grok-4
93.3%
#22
GPT-4.1
93.3%
#23
Grok-4-Thinking
93.3%
#24
Kimi-K2-Instruct
91.1%
#25
Granite-3.0
90.0%
#26
Llama-4-Maverick-17B
87.8%
#27
Gemini-1.5-Pro
86.7%
#28
Grok-3-Mini-Beta
84.4%
#29
GPT-4o-0513
81.1%
#30
Qwen-3
80.0%
#31
Qwen-3-Thinking
80.0%
#32
GPT-4-mini
78.9%
#33
Llama-3.1-405B
78.9%
#34
DeepSeek-R1
75.6%
#35
Grok-3-Beta
73.3%
#36
GPT-3.5-Turbo
67.8%
#37
Mistral-Large-2
64.4%
#38
DeepSeek-V3
62.2%
#39
OpenAI-O1-mini
51.1%
#40
Cohere-Command-R-Plus
43.3%