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 Haiku
100.0%
#2
Claude 3.5 Sonnet
100.0%
#3
Claude Haiku 4.5
100.0%
#4
Claude Haiku 4.5 (Thinking)
100.0%
#5
OpenAI o3 (High Effort)
100.0%
#6
OpenAI o3 Mini (Medium Effort)
100.0%
#7
OpenAI o3 Mini (High Effort)
100.0%
#8
Gemini 2.5 Pro
100.0%
#9
Gemini 2.5 Pro (Thinking)
100.0%
#10
OpenAI o3 (Medium Effort)
100.0%
#11
Claude 4.0 Sonnet
90.0%
#12
Claude 4.0 Sonnet (Thinking)
90.0%
#13
Kimi K2
86.7%
#14
Claude 3 Sonnet
86.7%
#15
Claude Sonnet 4.5
86.0%
#16
Claude Sonnet 4.5 (Thinking)
86.0%
#17
Claude 3.7 Sonnet
83.3%
#18
Claude 3.7 Sonnet (Thinking)
83.3%
#19
Gemini 2.0 Flash (Thinking)
80.0%
#20
Claude 3 Opus
80.0%
#21
Gemini 2.0 Flash Thinking (Jan 2025)
80.0%
#22
Gemini 2.0 Flash
80.0%
#23
Llama 4 Maverick 17B
77.3%
#24
GPT-4.1
76.7%
#25
OpenAI o1
68.0%
#26
Llama 3.1 405B
66.7%
#27
Granite 3.0
66.7%
#28
GPT-4o
66.7%
#29
GPT-4o Mini
66.7%
#30
GPT-3.5 Turbo
60.0%
#31
DeepSeek R1
60.0%
#32
GPT OSS 120B
60.0%
#33
Llama 3.3 70B
60.0%
#34
Gemini 3.1 Pro Preview
57.9%
#35
Claude 4.0 Opus (Thinking)
53.3%
#36
Claude 4.0 Opus
53.3%
#37
Gemini 1.5 Pro
52.0%
#38
OpenAI o4 Mini (Medium Effort)
50.0%
#39
DeepSeek V3
50.0%
#40
OpenAI o4 Mini (High Effort)
50.0%
#41
Mistral Large 2
43.3%
#42
Gemini 2.0 Pro
40.0%
#43
Cohere Command A
36.7%
#44
Cohere Command R+
36.7%
#45
GPT-5
28.3%
#46
Gemini 2.5 Flash
28.3%
#47
GPT-5 (Thinking)
28.3%
#48
Gemini 2.5 Flash (Thinking)
28.3%
#49
GPT-4.5
23.3%
#50
Qwen 3
23.3%
#51
Qwen 3 (Thinking)
23.3%
#52
Grok 3 Mini
18.3%
#53
Grok 3
18.3%
#54
OpenAI o1 Mini
18.3%
#55
Grok 3 (Thinking)
18.3%
#56
Grok 4
10.0%
#57
Grok 4 (Thinking)
10.0%