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 3.5 Sonnet
100.0%
#2
Claude 4.0 Sonnet
100.0%
#3
Claude Haiku 4.5 (Thinking)
100.0%
#4
Claude Haiku 4.5
100.0%
#5
Claude 4.0 Sonnet (Thinking)
100.0%
#6
Claude Sonnet 4.5
97.3%
#7
Claude Sonnet 4.5 (Thinking)
97.3%
#8
OpenAI o4 Mini (High Effort)
93.9%
#9
OpenAI o4 Mini (Medium Effort)
93.9%
#10
GPT-5 (Thinking)
93.3%
#11
GPT-4.5
93.3%
#12
GPT-5
93.3%
#13
Claude 3 Opus
93.3%
#14
Claude 3.5 Haiku
93.3%
#15
GPT OSS 120B
93.3%
#16
Claude 4.0 Opus (Thinking)
92.7%
#17
Claude 4.0 Opus
92.7%
#18
Kimi K2
90.0%
#19
OpenAI o3 (High Effort)
90.0%
#20
OpenAI o3 (Medium Effort)
90.0%
#21
Claude 3 Sonnet
88.0%
#22
OpenAI o3 Mini (High Effort)
86.7%
#23
OpenAI o3 Mini (Medium Effort)
86.7%
#24
Claude 3.7 Sonnet (Thinking)
86.7%
#25
Claude 3.7 Sonnet
86.7%
#26
Gemini 2.5 Pro (Thinking)
83.3%
#27
Gemini 2.5 Pro
83.3%
#28
Granite 3.0
80.0%
#29
GPT-4.1
80.0%
#30
OpenAI o1
75.3%
#31
Gemini 2.5 Flash
74.2%
#32
Gemini 2.5 Flash (Thinking)
74.2%
#33
Llama 4 Maverick 17B
70.0%
#34
Qwen 3 (Thinking)
65.0%
#35
Qwen 3
65.0%
#36
Gemini 2.0 Flash
63.3%
#37
Gemini 2.0 Pro
63.3%
#38
Gemini 2.0 Flash Thinking (Jan 2025)
63.3%
#39
Gemini 2.0 Flash (Thinking)
63.3%
#40
Llama 3.1 405B
63.3%
#41
GPT-4o
60.0%
#42
DeepSeek R1
51.7%
#43
Grok 3 Mini
50.0%
#44
OpenAI o1 Mini
40.0%
#45
Gemini 1.5 Pro
40.0%
#46
GPT-4o Mini
40.0%
#47
Gemini 3.1 Pro Preview
39.0%
#48
GPT-3.5 Turbo
33.3%
#49
Llama 3.3 70B
30.0%
#50
DeepSeek V3
23.3%
#51
Grok 3
20.0%
#52
Grok 3 (Thinking)
20.0%
#53
Cohere Command R+
16.7%
#54
Grok 4 (Thinking)
16.7%
#55
Grok 4
16.7%
#56
Mistral Large 2
13.3%
#57
Cohere Command A
10.0%