Skip to content

Benchmark

HELM Safety: HarmBench - LM Evaluated Safety score

Safety benchmarks examine risks in model behaviour, including how models respond to unsafe requests. Scores are percentages. Higher is better.

Models measured
69
Scores recorded
74
Sources
1
IQ weight
0.89

Score distribution

74.4%
Mean
76.8%
Median
19.7%
Std dev
46.9%
10th percentile
97.6%
90th percentile
23.3%
Min
23.3%100%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.89 for chat models, measured over 69 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1GPT OSS 120B· Stanford HELM specialized suites data docs100%
2GPT OSS 20b· Stanford HELM specialized suites data docs98.7%
3GPT-5 nano· Stanford HELM specialized suites data docs98.4%
4OpenAI o3· Stanford HELM specialized suites data docs98.4%
5Claude 3.5 Sonnet· Stanford HELM specialized suites data docs98.1%
6Claude Sonnet 4thinking, 10k· Stanford HELM specialized suites data docs98%
7GPT 5· Stanford HELM specialized suites data docs97.6%
8GPT 5.1· Stanford HELM specialized suites data docs97.6%
9Claude 3 Opus· Stanford HELM specialized suites data docs97.4%
10Kimi K2 Instruct· Stanford HELM specialized suites data docs97.4%
11GPT-5 mini· Stanford HELM specialized suites data docs97.1%
12o4 Mini· Stanford HELM specialized suites data docs97%
13o1· Stanford HELM specialized suites data docs96.3%
14Claude Haiku 4.5· Stanford HELM specialized suites data docs95.9%
15GPT-4.5 Preview· Stanford HELM specialized suites data docs95.8%
16Claude 3 Sonnet· Stanford HELM specialized suites data docs95.8%
17o3 Mini· Stanford HELM specialized suites data docs95.2%
18Claude Sonnet 4setting not named by the source· Stanford HELM specialized suites data docs94.1%
19Claude Sonnet 4.5· Stanford HELM specialized suites data docs92%
20Claude Opus 4thinking, 10k· Stanford HELM specialized suites data docs91.7%
21GPT 4.1· Stanford HELM specialized suites data docs91.7%
22Claude 3 Haiku· Stanford HELM specialized suites data docs91.3%
23GPT-4 Turbo· Stanford HELM specialized suites data docs89.8%
24Claude Opus 4setting not named by the source· Stanford HELM specialized suites data docs89.1%
25O1 Mini· Stanford HELM specialized suites data docs88.5%
26GPT-4.1 nano· Stanford HELM specialized suites data docs86.8%
27GPT-4.1 mini· Stanford HELM specialized suites data docs85.6%
28GPT-4o mini· Stanford HELM specialized suites data docs84.9%
29Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs84.6%
30Claude Sonnet 3.7· Stanford HELM specialized suites data docs84.3%
31Olmo 2 32B Instruct· Stanford HELM specialized suites data docs84.1%
32GPT 4o· Stanford HELM specialized suites data docs82.9%
33Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs81.2%
34gemini-1.5-flash-001· Stanford HELM specialized suites data docs80%
35gemini-1.5-pro-001· Stanford HELM specialized suites data docs79.9%
36Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs79.6%
37Qwen2 72B Instruct· Stanford HELM specialized suites data docs76.8%
38Palmyra X5· Stanford HELM specialized suites data docs73.8%
39Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs72.8%
40Llama3 8B Chat· Stanford HELM specialized suites data docs72.7%
41Gemini 3 Pro Preview· Stanford HELM specialized suites data docs72.5%
42Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs72.2%
43Granite 4.0 H Small· Stanford HELM specialized suites data docs71.1%
44GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs68.6%
45GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs68%
46Qwen2.5 7B Instruct Turbo· Stanford HELM specialized suites data docs67.7%
47Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs67%
48Gemini 2.0 Flash· Stanford HELM specialized suites data docs66.2%
49Llama 4 Maverick· Stanford HELM specialized suites data docs66.1%
50Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs65.4%

Showing 1–50 of 74 results.

Data credits: Stanford HELM (Stanford CRFM); scenario scores from Capabilities, Safety, Long Context, MedHELM, AIR-Bench, MMLU and Finance