Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-ChildSexualAbuse

A measure of model refusal for Child Harm (Level-1: Content Safety Risks, Level-2: Child Harm) related to child sexual abuse. Includes Level-4 risks like solicitation and child sexual abuse material (CSAM).
Source:

Model Performance

#1
Claude-3.5-Sonnet-1022
100.0%
#2
Granite-3.0
100.0%
#3
GPT-OSS-120B
96.7%
#4
Claude-3-Sonnet
96.7%
#5
Claude-4.0-Sonnet
93.3%
#6
GPT-4.5
93.3%
#7
Claude-4.0-Opus-Thinking
93.3%
#8
Claude-4.0-Opus
93.3%
#9
Llama-4-Maverick-17B
93.3%
#10
Claude-4.0-Sonnet-Thinking
93.3%
#11
Claude-3-Opus
91.7%
#12
OpenAI-O4-mini-medium
90.0%
#13
OpenAI-O1-1217
88.3%
#14
Kimi-K2-Instruct
86.7%
#15
OpenAI-O3-mini-medium
86.7%
#16
OpenAI-O3-medium
86.7%
#17
Claude-3.7-Sonnet-Thinking
83.3%
#18
Claude-3.7-Sonnet
83.3%
#19
Gemini-2.0-Pro-0121
83.3%
#20
Gemini-2.0-Flash
80.0%
#21
Gemini-2.5-Pro-0325
76.7%
#22
Gemini-1.5-Pro
76.7%
#23
GPT-4.1
76.7%
#24
Gemini-2.5-Pro-Thinking
76.7%
#25
Grok-3-Mini-Beta
71.7%
#26
Qwen-3
70.0%
#27
Gemini-2.5-Flash
70.0%
#28
Qwen-3-Thinking
70.0%
#29
GPT-4o-0513
66.7%
#30
GPT-4-mini
66.7%
#31
Grok-4
66.7%
#32
Grok-4-Thinking
66.7%
#33
Llama-3.1-405B
66.7%
#34
GPT-3.5-Turbo
63.3%
#35
DeepSeek-R1
60.0%
#36
Grok-3-Beta
58.3%
#37
OpenAI-O1-mini
53.3%
#38
DeepSeek-V3
50.0%
#39
Mistral-Large-2
43.3%
#40
Cohere-Command-R-Plus
36.7%