Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - Violence & Extremism

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
0.94

Score distribution

61.4%
Mean
63.3%
Median
20.6%
Std dev
29.7%
10th percentile
84.8%
90th percentile
16.1%
Min
16.1%94.3%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.94 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 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs94.3%
2GPT OSS 120B· Stanford HELM specialized suites data docs92.5%
3Claude Haiku 4.5· Stanford HELM specialized suites data docs91.1%
4GPT OSS 20b· Stanford HELM specialized suites data docs91.1%
5GPT-5 nano· Stanford HELM specialized suites data docs86.2%
6GPT 5· Stanford HELM specialized suites data docs86%
7Claude Sonnet 4.5· Stanford HELM specialized suites data docs85.8%
8Claude 3 Haiku· Stanford HELM specialized suites data docs84.8%
9GPT-5 mini· Stanford HELM specialized suites data docs84.8%
10Claude 3 Sonnet· Stanford HELM specialized suites data docs84.7%
11Claude Sonnet 4· Stanford HELM specialized suites data docs83.5%
12Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs83.1%
13OpenAI o3· Stanford HELM specialized suites data docs83.1%
14Claude 3 Opus· Stanford HELM specialized suites data docs82.8%
15Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs82.7%
16GPT 5.1· Stanford HELM specialized suites data docs82.7%
17Palmyra X5· Stanford HELM specialized suites data docs82.5%
18Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs81.6%
19Claude Opus 4· Stanford HELM specialized suites data docs79.5%
20Llama3 8B Chat· Stanford HELM specialized suites data docs79.3%
21o4 Mini· Stanford HELM specialized suites data docs79.3%
22Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs77.9%
23o1· Stanford HELM specialized suites data docs76.2%
24Claude Sonnet 3.7· Stanford HELM specialized suites data docs75.7%
25Granite 4.0 H Small· Stanford HELM specialized suites data docs74.3%
26Kimi K2 Instruct· Stanford HELM specialized suites data docs72.8%
27Llama 4 Maverick· Stanford HELM specialized suites data docs72.2%
28Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs70.5%
29GPT-4.5 Preview· Stanford HELM specialized suites data docs70.5%
30o3 Mini· Stanford HELM specialized suites data docs70.5%
31Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs69%
32GPT-4 Turbo· Stanford HELM specialized suites data docs68.2%
33gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs67.5%
34Gemini 2.0 Flash· Stanford HELM specialized suites data docs66.7%
35Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs65.6%
36GPT-4.1 nano· Stanford HELM specialized suites data docs64.5%
37Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs63.7%
38GPT 4osetting not named by the source· Stanford HELM specialized suites data docs63.3%
39GPT 4.1· Stanford HELM specialized suites data docs63.1%
40Gemini 3 Pro Preview· Stanford HELM specialized suites data docs63%
41GPT 4· Stanford HELM specialized suites data docs63%
42GPT-4o mini· Stanford HELM specialized suites data docs62.6%
43gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs62.5%
44Qwen2 72B Instruct· Stanford HELM specialized suites data docs61.9%
45GPT-4.1 mini· Stanford HELM specialized suites data docs60.6%
46Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs59.1%
47gemini-1.5-pro-002· Stanford HELM specialized suites data docs58.9%
48gemini-1.5-flash-002· Stanford HELM specialized suites data docs58.7%
49Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs56%
50GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs55.6%

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