Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - Hate/Toxicity

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
0.96

Score distribution

75.5%
Mean
77.7%
Median
15.6%
Std dev
50.6%
10th percentile
94%
90th percentile
37.3%
Min
37.3%98.7%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.96 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 docs98.7%
2Claude Haiku 4.5· Stanford HELM specialized suites data docs98.1%
3Claude Sonnet 4.5· Stanford HELM specialized suites data docs96.2%
4Claude Sonnet 4· Stanford HELM specialized suites data docs96%
5GPT OSS 20b· Stanford HELM specialized suites data docs95.6%
6Claude 3 Haiku· Stanford HELM specialized suites data docs95.1%
7Claude Opus 4· Stanford HELM specialized suites data docs94.2%
8GPT OSS 120B· Stanford HELM specialized suites data docs94%
9GPT 5.1· Stanford HELM specialized suites data docs93.6%
10Claude 3 Sonnet· Stanford HELM specialized suites data docs93.3%
11Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs92.8%
12Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs92.5%
13Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs92.1%
14Claude Sonnet 3.7· Stanford HELM specialized suites data docs91.7%
15GPT-5 nano· Stanford HELM specialized suites data docs91.4%
16Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs91.3%
17Claude 3 Opus· Stanford HELM specialized suites data docs91%
18Palmyra X5· Stanford HELM specialized suites data docs90.4%
19GPT 5· Stanford HELM specialized suites data docs89.4%
20GPT-4.5 Preview· Stanford HELM specialized suites data docs89.2%
21Granite 4.0 H Small· Stanford HELM specialized suites data docs89.1%
22GPT-5 mini· Stanford HELM specialized suites data docs87.9%
23Kimi K2 Instruct· Stanford HELM specialized suites data docs87.1%
24GPT-4 Turbo· Stanford HELM specialized suites data docs86%
25Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs83.3%
26Llama3 8B Chat· Stanford HELM specialized suites data docs82.2%
27Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs81.7%
28o3 Mini· Stanford HELM specialized suites data docs81.4%
29OpenAI o3· Stanford HELM specialized suites data docs81%
30Qwen2 72B Instruct· Stanford HELM specialized suites data docs80.7%
31Gemini 2.0 Flash· Stanford HELM specialized suites data docs80.2%
32GPT 4.1· Stanford HELM specialized suites data docs80%
33o1· Stanford HELM specialized suites data docs79.9%
34gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs79.4%
35Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs79.2%
36gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs78.6%
37Llama 4 Maverick· Stanford HELM specialized suites data docs78.3%
38GPT-4.1 mini· Stanford HELM specialized suites data docs77.7%
39o4 Mini· Stanford HELM specialized suites data docs77.7%
40GLM 4.5 Air· Stanford HELM specialized suites data docs77.6%
41Gemini 3 Pro Preview· Stanford HELM specialized suites data docs77.2%
42GPT 4osetting not named by the source· Stanford HELM specialized suites data docs76.6%
43GPT-4.1 nano· Stanford HELM specialized suites data docs75.6%
44GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs75.3%
45gemini-1.5-pro-002· Stanford HELM specialized suites data docs74.3%
46Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs74.1%
47Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs73.9%
48gemini-1.5-flash-002· Stanford HELM specialized suites data docs72.8%
49GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs72.3%
50GPT 4· Stanford HELM specialized suites data docs71.4%

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