Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-DepictingViolence

A measure of model refusal for Violence & Extremism (Level-1: Content Safety Risks, Level-2: Violence & Extremism) related to depicting violence. Includes Level-4 risks like bodily destruction, bodily mutilation, torture/abuse, animal abuse, and activities meant to kill.
Source:

Model Performance

#1
Claude 3.5 Haiku
100.0%
#2
OpenAI o3 (Medium Effort)
100.0%
#3
OpenAI o3 (High Effort)
100.0%
#4
Gemini 1.5 Pro
100.0%
#5
Claude 3 Sonnet
98.3%
#6
GPT-5
98.3%
#7
GPT-5 (Thinking)
98.3%
#8
OpenAI o1
96.7%
#9
Claude 4.0 Sonnet
94.2%
#10
Claude 4.0 Sonnet (Thinking)
94.2%
#11
Claude 4.0 Opus (Thinking)
93.3%
#12
Claude 4.0 Opus
93.3%
#13
Claude Haiku 4.5
92.8%
#14
Claude Haiku 4.5 (Thinking)
92.8%
#15
Claude 3.5 Sonnet
90.7%
#16
GPT-4.5
90.0%
#17
Gemini 3.1 Pro Preview
90.0%
#18
Gemini 2.5 Pro
88.3%
#19
Gemini 2.5 Pro (Thinking)
88.3%
#20
Gemini 2.5 Flash
86.7%
#21
Gemini 2.5 Flash (Thinking)
86.7%
#22
Llama 4 Maverick 17B
85.0%
#23
GPT OSS 120B
82.6%
#24
Granite 3.0
69.6%
#25
Claude Sonnet 4.5 (Thinking)
63.3%
#26
Claude Sonnet 4.5
63.3%
#27
Gemini 2.0 Pro
60.0%
#28
Claude 3 Opus
60.0%
#29
Claude 3.7 Sonnet (Thinking)
58.0%
#30
Claude 3.7 Sonnet
58.0%
#31
OpenAI o4 Mini (High Effort)
55.0%
#32
OpenAI o4 Mini (Medium Effort)
55.0%
#33
Grok 3 Mini
47.8%
#34
Gemini 2.0 Flash (Thinking)
42.7%
#35
Gemini 2.0 Flash
42.7%
#36
Gemini 2.0 Flash Thinking (Jan 2025)
42.7%
#37
Kimi K2
40.7%
#38
Grok 3
40.6%
#39
Grok 3 (Thinking)
40.6%
#40
OpenAI o1 Mini
40.0%
#41
Grok 4
37.7%
#42
Grok 4 (Thinking)
37.7%
#43
GPT-3.5 Turbo
37.0%
#44
Mistral Large 2
34.7%
#45
GPT-4o Mini
30.0%
#46
GPT-4o
28.0%
#47
Cohere Command R+
28.0%
#48
Llama 3.3 70B
28.0%
#49
Llama 3.1 405B
27.3%
#50
OpenAI o3 Mini (Medium Effort)
27.3%
#51
OpenAI o3 Mini (High Effort)
27.3%
#52
Cohere Command A
26.7%
#53
GPT-4.1
26.7%
#54
DeepSeek V3
24.0%
#55
DeepSeek R1
23.3%
#56
Qwen 3
23.3%
#57
Qwen 3 (Thinking)
23.3%