Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-Misdisinformation

A measure of model refusal for Deception (Level-1: Societal Risks, Level-2: Deception) related to mis/disinformation. Includes Level-4 risks like generating or promoting misinformation and fake online engagement (fake reviews, fake grassroots support).
Source:

Model Performance

#1
Claude-3-Opus
90.0%
#2
Claude-3.5-Sonnet-1022
90.0%
#3
Claude-4.0-Opus-Thinking
90.0%
#4
Claude-4.0-Opus
90.0%
#5
Kimi-K2-Instruct
90.0%
#6
OpenAI-O3-medium
90.0%
#7
GPT-OSS-120B
90.0%
#8
Claude-3.7-Sonnet
86.7%
#9
Claude-4.0-Sonnet
86.7%
#10
Claude-3.7-Sonnet-Thinking
86.7%
#11
Claude-3-Sonnet
86.7%
#12
Claude-4.0-Sonnet-Thinking
86.7%
#13
Granite-3.0
86.7%
#14
OpenAI-O1-1217
86.7%
#15
GPT-4.5
86.7%
#16
OpenAI-O4-mini-medium
86.7%
#17
GPT-4.1
80.0%
#18
Gemini-2.5-Pro-Thinking
73.3%
#19
Llama-4-Maverick-17B
73.3%
#20
Gemini-2.5-Pro-0325
73.3%
#21
OpenAI-O3-mini-medium
71.7%
#22
Gemini-2.0-Pro-0121
70.0%
#23
Gemini-2.5-Flash
66.7%
#24
Gemini-1.5-Pro
65.0%
#25
Gemini-2.0-Flash
63.3%
#26
Llama-3.1-405B
63.3%
#27
Qwen-3
60.0%
#28
Qwen-3-Thinking
60.0%
#29
GPT-4o-0513
60.0%
#30
DeepSeek-R1
51.7%
#31
GPT-4-mini
40.0%
#32
Grok-3-Mini-Beta
36.7%
#33
Grok-3-Beta
36.7%
#34
OpenAI-O1-mini
26.7%
#35
DeepSeek-V3
23.3%
#36
GPT-3.5-Turbo
23.3%
#37
Grok-4-Thinking
20.0%
#38
Grok-4
20.0%
#39
Cohere-Command-R-Plus
16.7%
#40
Mistral-Large-2
13.3%