WHBench - Harm Rate: leaderboard

Safety failures on WHBench (2026), the clinician-designed women's health benchmark of 47 open-ended scenarios: the share of a model's answers that fail either the unsafe-commission or the unsafe-omission criterion of the 23-item grading rubric, judged by Claude Sonnet 4.6 against clinician reference answers. Percent of responses; lower is better.

Metric: Harm rate (%): share of responses failing either the unsafe-commission or the unsafe-omission safety criterion on the 47 expert-written WHBench scenarios in ten women's health topics (fertility, hormonal health, pregnancy, PCOS, contraception and others), three runs per question at temperature 0 with a 4,096-token cap, each response graded by a Claude Sonnet 4.6 judge against 4-6 clinician reference answers and a 23-criterion rubric with asymmetric safety penalties; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 22 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.612.8#113
2Claude Sonnet 4.627#173
3Mistral Large 330.5#832
4Gemini 3 Flash (Preview)32.6#152
5Grok 333.3#611
6Grok 437.1#319
7O338.6#237
8DeepSeek V3.244#419
9GPT-5.447.5#135
10DeepSeek R147.5#482
11Grok 3 Mini53.9#565
12Claude Opus 456#326
13GPT-4.161#505
14Claude Sonnet 468.1#410
15Gemini 2.5 Flash73.8#463

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=whbench-harm-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.