Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #44: Illegal services/exploitation

Safety benchmarks examine risks in model behaviour, including how models respond to unsafe requests. Scores are percentages. Higher is better.

Models measured
70
Scores recorded
75
Sources
1
IQ weight
1.51

Score distribution

72%
Mean
82.2%
Median
26.5%
Std dev
31.1%
10th percentile
98.9%
90th percentile
2.2%
Min
2.2%100%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.51 for chat models, measured over 70 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Claude 3 Haiku· Stanford HELM specialized suites data docs100%
2Claude 3 Sonnet· Stanford HELM specialized suites data docs100%
3Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs100%
4Claude Haiku 4.5· Stanford HELM specialized suites data docs100%
5Claude Sonnet 4.5· Stanford HELM specialized suites data docs100%
6GPT OSS 20b· Stanford HELM specialized suites data docs100%
7Llama3 8B Chat· Stanford HELM specialized suites data docs100%
8Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs98.9%
9Claude 3 Opus· Stanford HELM specialized suites data docs97.8%
10GPT 5· Stanford HELM specialized suites data docs97.8%
11GPT OSS 120B· Stanford HELM specialized suites data docs97.8%
12GPT-5 mini· Stanford HELM specialized suites data docs97.8%
13Claude Sonnet 4· Stanford HELM specialized suites data docs95.6%
14GPT-5 nano· Stanford HELM specialized suites data docs95.6%
15Granite 4.0 H Small· Stanford HELM specialized suites data docs95.6%
16o4 Mini· Stanford HELM specialized suites data docs95.6%
17Palmyra X5· Stanford HELM specialized suites data docs95.6%
18Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs95.6%
19GPT-4.5 Preview· Stanford HELM specialized suites data docs93.3%
20Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs93.3%
21Llama 4 Maverick· Stanford HELM specialized suites data docs93.3%
22o3 Mini· Stanford HELM specialized suites data docs93.3%
23Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs93.3%
24Claude Sonnet 3.7· Stanford HELM specialized suites data docs91.1%
25Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs91.1%
26GPT 5.1· Stanford HELM specialized suites data docs91.1%
27o1· Stanford HELM specialized suites data docs91.1%
28OpenAI o3· Stanford HELM specialized suites data docs91.1%
29Claude Opus 4· Stanford HELM specialized suites data docs87.8%
30Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs86.7%
31gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs86.7%
32GPT-4 Turbo· Stanford HELM specialized suites data docs86.7%
33Kimi K2 Instruct· Stanford HELM specialized suites data docs86.7%
34Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs86.7%
35Gemini 2.0 Flash· Stanford HELM specialized suites data docs84.4%
36Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs84.4%
37Gemini 3 Pro Preview· Stanford HELM specialized suites data docs84.4%
38GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs82.2%
39GPT-4.1 nano· Stanford HELM specialized suites data docs82.2%
40GPT 4· Stanford HELM specialized suites data docs76.7%
41GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs73.3%
42GPT-4o mini· Stanford HELM specialized suites data docs73.3%
43gemini-1.5-flash-002· Stanford HELM specialized suites data docs71.1%
44gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs71.1%
45Qwen2 72B Instruct· Stanford HELM specialized suites data docs71.1%
46Yi 34B Chat· Stanford HELM specialized suites data docs67.8%
47gemini-1.5-pro-002· Stanford HELM specialized suites data docs66.7%
48GLM 4.5 Air· Stanford HELM specialized suites data docs66.7%
49GPT 4.1· Stanford HELM specialized suites data docs66.7%
50GPT 4osetting not named by the source· Stanford HELM specialized suites data docs66.7%

Showing 1–50 of 75 results.

Data credits: Stanford HELM (Stanford CRFM); scenario scores from Capabilities, Safety, Long Context, MedHELM, AIR-Bench, MMLU and Finance