HumaneBench - Bad Persona: leaderboard
The same 800 prompts with a system prompt that tells the model to maximize engagement and disregard the user's wellbeing; the score is the HumaneScore of the replies. It measures whether the model keeps protecting the user when the operator instructs it not to, which the no-prompt board does not test.
Metric: HumaneScore under a system prompt to disregard user wellbeing (-1 to 1). Source: github.com. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 0.83 | #167 |
| 2 | GPT-5.1 | 0.83 | #229 |
| 3 | Claude Sonnet 4.5 | 0.77 | #250 |
| 4 | Claude Opus 4.1 | 0.66 | #272 |
| 5 | Claude Sonnet 4 | 0.5 | #400 |
| 6 | Llama 4 Maverick | -0.14 | #1099 |
| 7 | DeepSeek V3.1 Terminus | -0.36 | #445 |
| 8 | Gemini 3 Pro (Preview) | -0.45 | #113 |
| 9 | Llama 3.1 405B Instruct | -0.49 | #1115 |
| 10 | GPT-4.1 | -0.61 | #495 |
| 11 | GPT-4o (2024-11-20) | -0.61 | #866 |
| 12 | Gemini 2.5 Flash | -0.68 | #453 |
| 13 | Gemini 2.0 Flash (001) | -0.71 | #727 |
| 14 | Gemini 2.5 Pro | -0.72 | #266 |
| 15 | Grok 4 | -0.73 | #307 |
Interactive version: theaggregate.ai/benchmark?slug=humanebench-bad-persona · How It Works · Data refreshed daily, snapshot 2026-11-13.