Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #26: Deterring Democratic Participation

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
0.80

Score distribution

58.9%
Mean
56.7%
Median
25.2%
Std dev
23.3%
10th percentile
95%
90th percentile
10%
Min
10%100%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.80 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 docs100%
2Claude Haiku 4.5· Stanford HELM specialized suites data docs100%
3Claude Sonnet 4.5· Stanford HELM specialized suites data docs98.3%
4Claude 3 Haiku· Stanford HELM specialized suites data docs96.7%
5Claude 3 Sonnet· Stanford HELM specialized suites data docs96.7%
6Claude 3 Opus· Stanford HELM specialized suites data docs95.8%
7Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs95%
8GPT OSS 20b· Stanford HELM specialized suites data docs95%
9Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs95%
10Claude Sonnet 4· Stanford HELM specialized suites data docs94.2%
11GPT OSS 120B· Stanford HELM specialized suites data docs92.5%
12Claude Opus 4· Stanford HELM specialized suites data docs90.8%
13GPT 5· Stanford HELM specialized suites data docs89.2%
14Claude Sonnet 3.7· Stanford HELM specialized suites data docs88.3%
15OpenAI o3· Stanford HELM specialized suites data docs87.5%
16Palmyra X5· Stanford HELM specialized suites data docs85.8%
17GPT-5 mini· Stanford HELM specialized suites data docs85%
18Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs84.2%
19GPT 5.1· Stanford HELM specialized suites data docs81.7%
20GPT-5 nano· Stanford HELM specialized suites data docs81.7%
21Kimi K2 Instruct· Stanford HELM specialized suites data docs81.7%
22Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs81.7%
23o4 Mini· Stanford HELM specialized suites data docs77.5%
24Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs72.5%
25o3 Mini· Stanford HELM specialized suites data docs71.7%
26o1· Stanford HELM specialized suites data docs70%
27Llama 4 Maverick· Stanford HELM specialized suites data docs67.5%
28Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs66.7%
29Gemini 2.0 Flash· Stanford HELM specialized suites data docs63.3%
30Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs63.3%
31GPT-4.5 Preview· Stanford HELM specialized suites data docs63.3%
32Granite 4.0 H Small· Stanford HELM specialized suites data docs63.3%
33gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs61.7%
34Llama3 8B Chat· Stanford HELM specialized suites data docs61.7%
35GLM 4.5 Air· Stanford HELM specialized suites data docs58.3%
36GPT-4 Turbo· Stanford HELM specialized suites data docs58.3%
37Qwen2 72B Instruct· Stanford HELM specialized suites data docs57.5%
38gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs56.7%
39Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs56.7%
40GPT 4.1· Stanford HELM specialized suites data docs55.8%
41GPT-4.1 nano· Stanford HELM specialized suites data docs55%
42Gemini 3 Pro Preview· Stanford HELM specialized suites data docs54.2%
43GPT-4.1 mini· Stanford HELM specialized suites data docs53.3%
44Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs51.7%
45gemini-1.5-pro-002· Stanford HELM specialized suites data docs51.7%
46GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs51.7%
47Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs51.7%
48GPT 4· Stanford HELM specialized suites data docs49.2%
49Yi 34B Chat· Stanford HELM specialized suites data docs48.3%
50GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs46.7%

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