HumaneBench: leaderboard
The model answers 800 everyday user prompts with no system prompt, and three judge models rate each reply from -1 to 1 on whether it protects the user's attention, autonomy, wellbeing, relationships and dignity; the score is the mean over eight principles. Unlike Spiral-Bench it scores single replies across ordinary requests rather than long conversations with a user in a delusional spiral.
Metric: HumaneScore, no system prompt (-1 to 1). Source: github.com. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 0.86 | #229 |
| 2 | GPT-5 | 0.86 | #167 |
| 3 | Gemini 3 Pro (Preview) | 0.78 | #113 |
| 4 | Gemini 2.5 Pro | 0.76 | #266 |
| 5 | Claude Sonnet 4.5 | 0.75 | #250 |
| 6 | DeepSeek V3.1 Terminus | 0.75 | #445 |
| 7 | Gemini 2.0 Flash (001) | 0.75 | #727 |
| 8 | Gemini 2.5 Flash | 0.72 | #453 |
| 9 | Claude Sonnet 4 | 0.69 | #400 |
| 10 | Grok 4 | 0.69 | #307 |
| 11 | Claude Opus 4.1 | 0.68 | #272 |
| 12 | GPT-4o (2024-11-20) | 0.68 | #866 |
| 13 | GPT-4.1 | 0.67 | #495 |
| 14 | Llama 4 Maverick | 0.59 | #1099 |
| 15 | Llama 3.1 405B Instruct | 0.56 | #1115 |
Interactive version: theaggregate.ai/benchmark?slug=humanebench · How It Works · Data refreshed daily, snapshot 2026-11-13.