Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - Refusal Rate

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
1.07

Score distribution

65%
Mean
66.2%
Median
17.4%
Std dev
40.8%
10th percentile
86.7%
90th percentile
25.4%
Min
25.4%93.2%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.07 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 docs93.2%
2Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs90.8%
3Claude Sonnet 4.5· Stanford HELM specialized suites data docs89.8%
4Claude Sonnet 4· Stanford HELM specialized suites data docs88.3%
5GPT OSS 120B· Stanford HELM specialized suites data docs88%
6GPT-5 nano· Stanford HELM specialized suites data docs87.8%
7GPT 5· Stanford HELM specialized suites data docs87.7%
8Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs86.7%
9GPT 5.1· Stanford HELM specialized suites data docs86.2%
10GPT OSS 20b· Stanford HELM specialized suites data docs86%
11Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs85.9%
12Claude Opus 4· Stanford HELM specialized suites data docs85.7%
13GPT-5 mini· Stanford HELM specialized suites data docs85.7%
14Claude 3 Sonnet· Stanford HELM specialized suites data docs84.7%
15OpenAI o3· Stanford HELM specialized suites data docs84.5%
16Claude 3 Opus· Stanford HELM specialized suites data docs84.4%
17Claude 3 Haiku· Stanford HELM specialized suites data docs82.7%
18Claude Sonnet 3.7· Stanford HELM specialized suites data docs81.8%
19o1· Stanford HELM specialized suites data docs80%
20Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs79%
21o4 Mini· Stanford HELM specialized suites data docs78.5%
22Palmyra X5· Stanford HELM specialized suites data docs78.2%
23Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs76%
24o3 Mini· Stanford HELM specialized suites data docs74.9%
25GPT-4.5 Preview· Stanford HELM specialized suites data docs74.1%
26Kimi K2 Instruct· Stanford HELM specialized suites data docs74.1%
27Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs73.6%
28Gemini 3 Pro Preview· Stanford HELM specialized suites data docs73.2%
29GPT-4 Turbo· Stanford HELM specialized suites data docs71.9%
30Granite 4.0 H Small· Stanford HELM specialized suites data docs71.6%
31Llama3 8B Chat· Stanford HELM specialized suites data docs70.9%
32gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs68.7%
33Llama 4 Maverick· Stanford HELM specialized suites data docs68.6%
34gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs68.4%
35Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs67.5%
36gemini-1.5-pro-002· Stanford HELM specialized suites data docs67.3%
37gemini-1.5-flash-002· Stanford HELM specialized suites data docs67.1%
38Gemini 2.0 Flash· Stanford HELM specialized suites data docs66.2%
39Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs65.8%
40GPT 4.1· Stanford HELM specialized suites data docs64.8%
41GPT 4· Stanford HELM specialized suites data docs64.2%
42GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs63.5%
43GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs63.1%
44GPT 4osetting not named by the source· Stanford HELM specialized suites data docs62.3%
45Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs62.3%
46Qwen2 72B Instruct· Stanford HELM specialized suites data docs62.1%
47GPT-4.1 nano· Stanford HELM specialized suites data docs61.5%
48GPT-4.1 mini· Stanford HELM specialized suites data docs60.4%
49Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs59%
50Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs58.6%

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