Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #5: Autonomous Unsafe Operation of Systems

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
0.92

Score distribution

51%
Mean
52.7%
Median
18.8%
Std dev
23.9%
10th percentile
75.5%
90th percentile
14.2%
Min
14.2%86.1%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.92 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 Haiku 4.5· Stanford HELM specialized suites data docs86.1%
2GPT-5 nano· Stanford HELM specialized suites data docs86.1%
3OpenAI o3· Stanford HELM specialized suites data docs81.2%
4Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs80.9%
5Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs79.4%
6GPT OSS 120B· Stanford HELM specialized suites data docs79.1%
7GPT OSS 20b· Stanford HELM specialized suites data docs76.1%
8GPT-5 mini· Stanford HELM specialized suites data docs75.5%
9GPT 5· Stanford HELM specialized suites data docs74.8%
10Claude Sonnet 4.5· Stanford HELM specialized suites data docs74.5%
11Claude 3 Opus· Stanford HELM specialized suites data docs72.4%
12Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs70.9%
13GPT 5.1· Stanford HELM specialized suites data docs70.6%
14Claude 3 Haiku· Stanford HELM specialized suites data docs69.1%
15Claude Opus 4· Stanford HELM specialized suites data docs68.8%
16o1· Stanford HELM specialized suites data docs68.8%
17gemini-1.5-flash-002· Stanford HELM specialized suites data docs68.2%
18o3 Mini· Stanford HELM specialized suites data docs67.3%
19Claude Sonnet 4· Stanford HELM specialized suites data docs67%
20Claude 3 Sonnet· Stanford HELM specialized suites data docs66.7%
21GPT 4· Stanford HELM specialized suites data docs65.2%
22GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs63.3%
23Palmyra X5· Stanford HELM specialized suites data docs61.2%
24GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs60.9%
25Gemini 3 Pro Preview· Stanford HELM specialized suites data docs60.3%
26gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs60%
27o4 Mini· Stanford HELM specialized suites data docs60%
28Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs59.7%
29gemini-1.5-pro-002· Stanford HELM specialized suites data docs58.5%
30Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs57.6%
31Claude Sonnet 3.7· Stanford HELM specialized suites data docs57.3%
32GPT-4 Turbo· Stanford HELM specialized suites data docs55.2%
33Granite 4.0 H Small· Stanford HELM specialized suites data docs55.2%
34GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs54.8%
35Llama 4 Maverick· Stanford HELM specialized suites data docs53.9%
36Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs53.9%
37Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs53.3%
38Gemini 2.0 Flash· Stanford HELM specialized suites data docs52.7%
39Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs52.4%
40gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs52.4%
41Llama3 8B Chat· Stanford HELM specialized suites data docs52.4%
42GPT-4.5 Preview· Stanford HELM specialized suites data docs49.4%
43Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs48.5%
44Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs48.2%
45GPT 4osetting not named by the source· Stanford HELM specialized suites data docs47.9%
46Kimi K2 Instruct· Stanford HELM specialized suites data docs46.7%
47Yi 34B Chat· Stanford HELM specialized suites data docs43%
48GPT-4.1 mini· Stanford HELM specialized suites data docs42.1%
49Grok 3 Mini· Stanford HELM specialized suites data docs41.8%
50DeepSeek LLM 67B Chat· Stanford HELM specialized suites data docs40.3%

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