Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #4: Automated Decision-Making

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
0.72

Score distribution

34.9%
Mean
30%
Median
19.2%
Std dev
12.7%
10th percentile
66.3%
90th percentile
9%
Min
9%81%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.72 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 docs81%
2Claude Sonnet 4· Stanford HELM specialized suites data docs73%
3GPT OSS 120B· Stanford HELM specialized suites data docs71.7%
4Claude Opus 4· Stanford HELM specialized suites data docs70.7%
5Claude Sonnet 4.5· Stanford HELM specialized suites data docs68.3%
6Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs66.7%
7Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs66.7%
8Claude Sonnet 3.7· Stanford HELM specialized suites data docs66.3%
9Claude 3 Sonnet· Stanford HELM specialized suites data docs65%
10gemini-1.5-flash-002· Stanford HELM specialized suites data docs65%
11OpenAI o3· Stanford HELM specialized suites data docs62%
12Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs59.3%
13GPT 5.1· Stanford HELM specialized suites data docs59.3%
14GPT 5· Stanford HELM specialized suites data docs58%
15Claude 3 Opus· Stanford HELM specialized suites data docs57.3%
16gemini-1.5-pro-002· Stanford HELM specialized suites data docs56.3%
17GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs52.7%
18GPT-5 nano· Stanford HELM specialized suites data docs51.7%
19GPT OSS 20b· Stanford HELM specialized suites data docs51%
20GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs50.3%
21GPT 4· Stanford HELM specialized suites data docs48.7%
22gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs46.7%
23GPT-5 mini· Stanford HELM specialized suites data docs43.7%
24GPT-4 Turbo· Stanford HELM specialized suites data docs42%
25Palmyra X5· Stanford HELM specialized suites data docs42%
26Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs38.3%
27Gemini 3 Pro Preview· Stanford HELM specialized suites data docs36.7%
28DeepSeek LLM 67B Chat· Stanford HELM specialized suites data docs35.7%
29Yi 34B Chat· Stanford HELM specialized suites data docs35.3%
30Claude 3 Haiku· Stanford HELM specialized suites data docs34%
31Granite 4.0 H Small· Stanford HELM specialized suites data docs34%
32o1· Stanford HELM specialized suites data docs33%
33gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs31.7%
34Gemini 2.0 Flash· Stanford HELM specialized suites data docs31.3%
35o4 Mini· Stanford HELM specialized suites data docs31.3%
36GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs31%
37Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs30.7%
38Mixtral 8x22B Instruct v0.1· Stanford HELM specialized suites data docs30%
39Qwen1.5-72B-Chat· Stanford HELM specialized suites data docs29.3%
40O1 Mini· Stanford HELM specialized suites data docs28.3%
41GPT 4osetting not named by the source· Stanford HELM specialized suites data docs27.7%
42Mistral (7B) Instruct v0.3· Stanford HELM specialized suites data docs27.7%
43Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs27%
44GPT-4.5 Preview· Stanford HELM specialized suites data docs25.3%
45Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs25%
46Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs24.7%
47Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs24.7%
48Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs24.3%
49Llama 4 Maverick· Stanford HELM specialized suites data docs23%
50GPT-4o mini· Stanford HELM specialized suites data docs22.3%

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