Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - Child 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
1.08

Score distribution

57.8%
Mean
58.1%
Median
22.3%
Std dev
23.8%
10th percentile
85.7%
90th percentile
6.7%
Min
6.7%92.4%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.08 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 docs92.4%
2Claude Sonnet 4.5· Stanford HELM specialized suites data docs90%
3Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs89%
4Claude Sonnet 4· Stanford HELM specialized suites data docs88.6%
5Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs87.1%
6Claude 3 Sonnet· Stanford HELM specialized suites data docs86.7%
7Claude Opus 4· Stanford HELM specialized suites data docs86.7%
8Palmyra X5· Stanford HELM specialized suites data docs85.7%
9Claude 3 Haiku· Stanford HELM specialized suites data docs84.8%
10GPT OSS 20b· Stanford HELM specialized suites data docs82.9%
11GPT OSS 120B· Stanford HELM specialized suites data docs81.9%
12Llama 4 Maverick· Stanford HELM specialized suites data docs81.9%
13GPT-5 nano· Stanford HELM specialized suites data docs81%
14Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs81%
15Claude 3 Opus· Stanford HELM specialized suites data docs80.5%
16GPT 5· Stanford HELM specialized suites data docs80%
17GPT 5.1· Stanford HELM specialized suites data docs79%
18Claude Sonnet 3.7· Stanford HELM specialized suites data docs78.1%
19GPT-4.5 Preview· Stanford HELM specialized suites data docs77.1%
20o4 Mini· Stanford HELM specialized suites data docs77.1%
21OpenAI o3· Stanford HELM specialized suites data docs77.1%
22GPT-5 mini· Stanford HELM specialized suites data docs76.2%
23o3 Mini· Stanford HELM specialized suites data docs74.8%
24Gemini 3 Pro Preview· Stanford HELM specialized suites data docs74.3%
25o1· Stanford HELM specialized suites data docs73.8%
26Llama3 8B Chat· Stanford HELM specialized suites data docs73.3%
27Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs71.4%
28Kimi K2 Instruct· Stanford HELM specialized suites data docs71.4%
29Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs71.4%
30Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs66.7%
31Granite 4.0 H Small· Stanford HELM specialized suites data docs66.7%
32gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs64.8%
33GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs62.9%
34GPT-4 Turbo· Stanford HELM specialized suites data docs61.4%
35gemini-1.5-pro-002· Stanford HELM specialized suites data docs59%
36Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs58.1%
37gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs58.1%
38GPT 4.1· Stanford HELM specialized suites data docs58.1%
39GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs58.1%
40Gemini 2.0 Flash· Stanford HELM specialized suites data docs57.6%
41Llama 4 Scout· Stanford HELM specialized suites data docs57.1%
42Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs56.7%
43Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs55.2%
44gemini-1.5-flash-002· Stanford HELM specialized suites data docs54.8%
45GPT 4· Stanford HELM specialized suites data docs53.3%
46Grok 4· Stanford HELM specialized suites data docs53.3%
47GPT-4.1 nano· Stanford HELM specialized suites data docs52.9%
48Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs52.4%
49GPT-4.1 mini· Stanford HELM specialized suites data docs51.4%
50Qwen2 72B Instruct· Stanford HELM specialized suites data docs51.4%

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