Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-Harassment

A measure of model refusal for Hate/Toxicity (Level-1: Content Safety Risks, Level-2: Hate/Toxicity) related to harassment. Includes Level-4 risks like bullying, threats, intimidation, shaming, humiliation, insults/personal attacks, abuse, provoking, trolling, doxxing, and cursing.
Source:

Model Performance

#1
GPT-5 (Thinking)
100.0%
#2
GPT-5
100.0%
#3
GPT-3.5 Turbo
100.0%
#4
OpenAI o3 (Medium Effort)
100.0%
#5
OpenAI o1
100.0%
#6
Qwen 3 (Thinking)
100.0%
#7
Qwen 3
100.0%
#8
OpenAI o3 (High Effort)
100.0%
#9
GPT OSS 120B
98.9%
#10
Claude 3.5 Sonnet
98.7%
#11
Claude 3 Sonnet
96.7%
#12
Claude Sonnet 4.5
93.8%
#13
Claude Sonnet 4.5 (Thinking)
93.8%
#14
Claude Haiku 4.5
90.0%
#15
Claude 3.5 Haiku
90.0%
#16
Claude Haiku 4.5 (Thinking)
90.0%
#17
Gemini 2.5 Pro (Thinking)
88.3%
#18
Gemini 2.5 Pro
88.3%
#19
Gemini 2.5 Flash
85.0%
#20
Gemini 2.5 Flash (Thinking)
85.0%
#21
Claude 3 Opus
84.7%
#22
Claude 4.0 Opus (Thinking)
83.3%
#23
Claude 4.0 Sonnet (Thinking)
83.3%
#24
Claude 4.0 Opus
83.3%
#25
Claude 4.0 Sonnet
83.3%
#26
Llama 4 Maverick 17B
83.3%
#27
Granite 3.0
83.3%
#28
Claude 3.7 Sonnet
82.5%
#29
Claude 3.7 Sonnet (Thinking)
82.5%
#30
GPT-4.5
82.2%
#31
Kimi K2
80.9%
#32
Gemini 3.1 Pro Preview
80.0%
#33
Gemini 1.5 Pro
73.3%
#34
OpenAI o3 Mini (High Effort)
70.0%
#35
OpenAI o3 Mini (Medium Effort)
70.0%
#36
GPT-4.1
64.0%
#37
Gemini 2.0 Flash (Thinking)
63.2%
#38
Gemini 2.0 Flash Thinking (Jan 2025)
63.2%
#39
Gemini 2.0 Flash
63.2%
#40
Grok 3
59.4%
#41
Grok 3 (Thinking)
59.4%
#42
GPT-4o
54.8%
#43
Llama 3.1 405B
50.3%
#44
OpenAI o4 Mini (High Effort)
50.0%
#45
OpenAI o4 Mini (Medium Effort)
50.0%
#46
GPT-4o Mini
48.9%
#47
Gemini 2.0 Pro
48.7%
#48
Grok 3 Mini
47.8%
#49
DeepSeek R1
43.5%
#50
OpenAI o1 Mini
40.0%
#51
Llama 3.3 70B
36.0%
#52
DeepSeek V3
35.8%
#53
Cohere Command A
26.6%
#54
Mistral Large 2
26.3%
#55
Cohere Command R+
20.4%
#56
Grok 4
18.0%
#57
Grok 4 (Thinking)
18.0%