PrincipalBench: leaderboard
Metric: Aggregate Harm Rate (%; 36 items x 3 arms, mean of 5 seeds). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 5.5 |
| 2 | Mistral Large 2 (Nov) Instruct (2411) | 11 |
| 3 | Gemini 3.1 Flash Lite (Preview) | 12 |
| 4 | DeepSeek V3.1 | 12.3 |
| 5 | Qwen 3 32B | 16.5 |
| 6 | Claude Opus 4.1 (20250805) | 18.1 |
| 7 | Llama 3.1 70B Instruct | 19.2 |
| 8 | Gemini 3 Flash (Preview) | 19.4 |
| 9 | Claude Sonnet 4.5 | 19.5 |
| 10 | GLM-4.6 | 46 |
| 11 | GPT-5 Mini | 53.6 |
| 12 | GPT-5 | 71.1 |
| 13 | Qwen 3.5 27B | 75.3 |
Interactive version: theaggregate.ai/benchmark?slug=principalbench · How It Works · Data refreshed daily, snapshot 2026-09-25.