MedHarm - Safe Helpfulness: leaderboard
MedHarm (2026) is a medical safety benchmark of 1,100 medically framed harmful requests in ten high-risk categories, such as covert poisoning, lethal dosing or weaponized pathogens, which a model should decline or redirect; a grading model scores each response. This board is the mean score in percent for responses that avoid harmful specifics while still offering appropriate clinical guidance or redirection; higher is better.
Metric: Safe helpfulness (%): mean judge score for responses that avoid actionable harm while still giving clinically appropriate general guidance or redirection; MedHarm: 1,100 high-risk medical queries in ten categories of 110 (covert poisoning, lethal dosing, lethal drug interactions, euthanasia, illegal organ harvesting and live anesthesia, drug abuse, medicalized chemical and biological weapons, date-rape drugs, fetal harm and more), all requests a model should decline or redirect; greedy decoding, 512 new tokens; GPT-5-mini judge (91% agreement with human labels on unsafe responses); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.3 Instant | 99.7 | #130 |
| 2 | Llama 3.1 8B Instruct | 96.2 | #4246 |
| 3 | GPT-5.5 | 95.5 | #49 |
| 4 | DeepSeek V4 Pro | 92 | #188 |
| 5 | Grok 4.3 | 91.8 | #333 |
| 6 | Qwen 2.5 7B Instruct | 81.3 | #2564 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medharm-safe-helpfulness · How It Works · Data refreshed daily, snapshot 2026-11-13.