Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-AcademicDishonesty

A measure of model refusal for Deception (Level-1: Societal Risks, Level-2: Deception) related to academic dishonesty. Includes Level-4 risks like plagiarism and promoting academic dishonesty.
Source:

Model Performance

#1
Claude-3-Opus
100.0%
#2
Claude-3-Sonnet
100.0%
#3
Claude-3.5-Sonnet-1022
100.0%
#4
GPT-OSS-120B
100.0%
#5
OpenAI-O4-mini-medium
100.0%
#6
OpenAI-O3-medium
100.0%
#7
Claude-3.7-Sonnet-Thinking
96.7%
#8
Claude-3.7-Sonnet
96.7%
#9
OpenAI-O3-mini-medium
96.7%
#10
Granite-3.0
96.7%
#11
Claude-4.0-Sonnet
96.7%
#12
Claude-4.0-Sonnet-Thinking
96.7%
#13
OpenAI-O1-1217
96.7%
#14
Claude-4.0-Opus
93.3%
#15
Claude-4.0-Opus-Thinking
93.3%
#16
Gemini-2.5-Pro-Thinking
93.3%
#17
Gemini-2.5-Pro-0325
93.3%
#18
GPT-4.5
93.3%
#19
Kimi-K2-Instruct
93.3%
#20
GPT-4.1
83.3%
#21
Gemini-1.5-Pro
80.0%
#22
Llama-4-Maverick-17B
80.0%
#23
Gemini-2.0-Flash
76.7%
#24
Gemini-2.0-Pro-0121
76.7%
#25
Qwen-3-Thinking
76.7%
#26
Qwen-3
76.7%
#27
GPT-4-mini
73.3%
#28
Gemini-2.5-Flash
66.7%
#29
DeepSeek-R1
60.0%
#30
GPT-4o-0513
60.0%
#31
Llama-3.1-405B
58.3%
#32
GPT-3.5-Turbo
53.3%
#33
Grok-3-Mini-Beta
53.3%
#34
Grok-3-Beta
51.7%
#35
OpenAI-O1-mini
46.7%
#36
Mistral-Large-2
38.3%
#37
Grok-4-Thinking
36.7%
#38
Grok-4
36.7%
#39
DeepSeek-V3
30.0%
#40
Cohere-Command-R-Plus
15.0%