Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - Manipulation

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
0.91

Score distribution

63.9%
Mean
63.3%
Median
25%
Std dev
27.3%
10th percentile
94.7%
90th percentile
8.7%
Min
8.7%99.3%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.91 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
1GPT 5· Stanford HELM specialized suites data docs99.3%
2GPT-5 nano· Stanford HELM specialized suites data docs99.3%
3Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs96.7%
4GPT OSS 20b· Stanford HELM specialized suites data docs96%
5Claude Haiku 4.5· Stanford HELM specialized suites data docs95.3%
6GPT-5 mini· Stanford HELM specialized suites data docs95.3%
7Claude 3 Haiku· Stanford HELM specialized suites data docs94.7%
8Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs94.7%
9Claude Sonnet 4.5· Stanford HELM specialized suites data docs94.7%
10GPT OSS 120B· Stanford HELM specialized suites data docs94.7%
11o1· Stanford HELM specialized suites data docs94.7%
12OpenAI o3· Stanford HELM specialized suites data docs94.7%
13GPT 5.1· Stanford HELM specialized suites data docs93.3%
14Claude Sonnet 4· Stanford HELM specialized suites data docs92.7%
15Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs92%
16Claude Opus 4· Stanford HELM specialized suites data docs91.3%
17Gemini 3 Pro Preview· Stanford HELM specialized suites data docs90.7%
18Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs90.7%
19Claude 3 Opus· Stanford HELM specialized suites data docs90%
20o4 Mini· Stanford HELM specialized suites data docs90%
21Kimi K2 Instruct· Stanford HELM specialized suites data docs88.7%
22Claude 3 Sonnet· Stanford HELM specialized suites data docs88%
23Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs88%
24Claude Sonnet 3.7· Stanford HELM specialized suites data docs86.7%
25o3 Mini· Stanford HELM specialized suites data docs80.7%
26Palmyra X5· Stanford HELM specialized suites data docs77.3%
27GPT-4.5 Preview· Stanford HELM specialized suites data docs74.7%
28Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs74%
29gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs71.3%
30GPT-4 Turbo· Stanford HELM specialized suites data docs70%
31Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs69.3%
32Llama 4 Maverick· Stanford HELM specialized suites data docs68.7%
33Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs67.3%
34Llama3 8B Chat· Stanford HELM specialized suites data docs66%
35Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs65.3%
36GPT 4.1· Stanford HELM specialized suites data docs64.7%
37Granite 4.0 H Small· Stanford HELM specialized suites data docs64.7%
38gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs63.3%
39GPT-4.1 nano· Stanford HELM specialized suites data docs63.3%
40Gemini 2.0 Flash· Stanford HELM specialized suites data docs62.7%
41Qwen2 72B Instruct· Stanford HELM specialized suites data docs60%
42GLM 4.5 Air· Stanford HELM specialized suites data docs59.3%
43GPT 4· Stanford HELM specialized suites data docs59.3%
44GPT-4.1 mini· Stanford HELM specialized suites data docs58%
45GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs57.3%
46GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs56.7%
47Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs56.7%
48gemini-1.5-flash-002· Stanford HELM specialized suites data docs56%
49Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs56%
50gemini-1.5-pro-002· Stanford HELM specialized suites data docs55.3%

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