Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-DisempoweringWorkers

A measure of model refusal for Economic Harm (Level-1: Societal Risks, Level-2: Economic Harm) related to disempowering workers. Includes Level-4 risks like undermining workers' rights, worsening job quality, encouraging undue worker surveillance, and causing harmful labor-force disruptions.
Source:

Model Performance

#1
Claude 3 Sonnet
100.0%
#2
Claude 4.0 Sonnet
100.0%
#3
Claude 3.5 Sonnet
100.0%
#4
Claude Sonnet 4.5
100.0%
#5
Claude 4.0 Sonnet (Thinking)
100.0%
#6
GPT-4.5
100.0%
#7
Claude Sonnet 4.5 (Thinking)
100.0%
#8
Granite 3.0
100.0%
#9
GPT OSS 120B
100.0%
#10
Gemini 2.5 Pro
100.0%
#11
Gemini 2.5 Pro (Thinking)
100.0%
#12
Grok 3 Mini
100.0%
#13
Claude Haiku 4.5
98.9%
#14
Claude Haiku 4.5 (Thinking)
98.9%
#15
Claude 3.5 Haiku
96.7%
#16
Claude 4.0 Opus
96.7%
#17
Claude 4.0 Opus (Thinking)
96.7%
#18
OpenAI o1
96.7%
#19
OpenAI o3 (High Effort)
96.0%
#20
OpenAI o3 (Medium Effort)
96.0%
#21
Claude 3 Opus
95.8%
#22
Claude 3.7 Sonnet (Thinking)
95.0%
#23
Claude 3.7 Sonnet
95.0%
#24
Llama 4 Maverick 17B
93.3%
#25
Gemini 2.0 Pro
90.0%
#26
Kimi K2
88.3%
#27
OpenAI o4 Mini (Medium Effort)
88.0%
#28
OpenAI o4 Mini (High Effort)
88.0%
#29
GPT-5
86.7%
#30
GPT-5 (Thinking)
86.7%
#31
OpenAI o3 Mini (Medium Effort)
85.8%
#32
OpenAI o3 Mini (High Effort)
85.8%
#33
Gemini 2.5 Flash
80.0%
#34
Gemini 2.5 Flash (Thinking)
80.0%
#35
Gemini 1.5 Pro
76.7%
#36
Gemini 2.0 Flash Thinking (Jan 2025)
75.0%
#37
Gemini 2.0 Flash
75.0%
#38
Gemini 2.0 Flash (Thinking)
75.0%
#39
Qwen 3
73.3%
#40
Qwen 3 (Thinking)
73.3%
#41
Grok 3 (Thinking)
65.0%
#42
Grok 3
65.0%
#43
GPT-4.1
63.3%
#44
Llama 3.1 405B
60.0%
#45
GPT-3.5 Turbo
50.0%
#46
GPT-4o Mini
43.3%
#47
DeepSeek R1
37.5%
#48
GPT-4o
36.7%
#49
Llama 3.3 70B
30.0%
#50
OpenAI o1 Mini
28.3%
#51
Grok 4 (Thinking)
26.7%
#52
Grok 4
26.7%
#53
Gemini 3.1 Pro Preview
25.0%
#54
DeepSeek V3
24.2%
#55
Mistral Large 2
20.0%
#56
Cohere Command A
18.3%
#57
Cohere Command R+
13.3%