Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - Self-harm

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
1.99

Score distribution

83.4%
Mean
90%
Median
17.6%
Std dev
56.7%
10th percentile
100%
90th percentile
27.8%
Min
27.8%100%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.99 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 Sonnet· Stanford HELM specialized suites data docs100%
2Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs100%
3Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs100%
4Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs100%
5Gemini 3 Pro Preview· Stanford HELM specialized suites data docs100%
6GPT 5· Stanford HELM specialized suites data docs100%
7GPT OSS 120B· Stanford HELM specialized suites data docs100%
8GPT OSS 20b· Stanford HELM specialized suites data docs100%
9GPT-5 nano· Stanford HELM specialized suites data docs100%
10o1· Stanford HELM specialized suites data docs100%
11o4 Mini· Stanford HELM specialized suites data docs100%
12OpenAI o3· Stanford HELM specialized suites data docs100%
13Claude 3 Haiku· Stanford HELM specialized suites data docs97.8%
14Claude 3 Opus· Stanford HELM specialized suites data docs97.8%
15Claude Sonnet 3.7· Stanford HELM specialized suites data docs97.8%
16Claude Sonnet 4· Stanford HELM specialized suites data docs97.8%
17GPT 5.1· Stanford HELM specialized suites data docs97.8%
18GPT-4.5 Preview· Stanford HELM specialized suites data docs97.8%
19GPT-5 mini· Stanford HELM specialized suites data docs97.8%
20Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs97.8%
21Claude Haiku 4.5· Stanford HELM specialized suites data docs96.7%
22Claude Opus 4· Stanford HELM specialized suites data docs96.7%
23Claude Sonnet 4.5· Stanford HELM specialized suites data docs95.6%
24Gemini 2.0 Flash· Stanford HELM specialized suites data docs95.6%
25gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs95.6%
26Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs95.6%
27Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs93.3%
28gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs93.3%
29GPT 4.1· Stanford HELM specialized suites data docs93.3%
30Grok 4· Stanford HELM specialized suites data docs93.3%
31o3 Mini· Stanford HELM specialized suites data docs93.3%
32Palmyra X5· Stanford HELM specialized suites data docs92.2%
33Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs91.1%
34Kimi K2 Instruct· Stanford HELM specialized suites data docs91.1%
35GPT 4osetting not named by the source· Stanford HELM specialized suites data docs90%
36GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs90%
37GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs90%
38Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs90%
39Llama3 8B Chat· Stanford HELM specialized suites data docs90%
40GLM 4.5 Air· Stanford HELM specialized suites data docs88.9%
41Granite 4.0 H Small· Stanford HELM specialized suites data docs88.9%
42GPT 4· Stanford HELM specialized suites data docs87.8%
43Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs87.8%
44Llama 4 Maverick· Stanford HELM specialized suites data docs87.8%
45gemini-1.5-flash-002· Stanford HELM specialized suites data docs86.7%
46gemini-1.5-pro-002· Stanford HELM specialized suites data docs86.7%
47GPT-4 Turbo· Stanford HELM specialized suites data docs84.4%
48GPT-4.1 nano· Stanford HELM specialized suites data docs84.4%
49Grok 3 Mini· Stanford HELM specialized suites data docs84.4%
50Qwen2 72B Instruct· Stanford HELM specialized suites data docs82.2%

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