HumanAgencyBench: leaderboard
HumanAgencyBench (HAB): 3,000 simulated user prompts, 500 per dimension, testing whether an AI assistant supports human agency. Each response is scored 0-10 by an o3 judge against dimension rubrics; this series is the authors' overall score, the mean of the six dimension means, on a 0-100 scale. Higher is better.
Metric: HAB score (0-100, mean of six dimension scores). Source: github.com. Saturation forecast: Around December 2027. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 71.3 | #94 |
| 2 | Claude Opus 4.5 (20251101) | 68.55 | #81 |
| 3 | Claude Opus 4.7 | 65.1 | #51 |
| 4 | Claude Opus 4.6 | 63.23 | #67 |
| 5 | Claude Haiku 4.5 (20251001) | 60.88 | #264 |
| 6 | Claude 3.5 Sonnet (20241022) | 54.57 | #302 |
| 7 | Claude Opus 4.1 (20250805) | 53.35 | #139 |
| 8 | GPT-5 (High) | 52.66 | #105 (GPT-5) |
| 9 | GPT-5 | 48.48 | #105 |
| 10 | Claude Sonnet 4 (20250514) | 45.9 | #223 |
| 11 | Claude 3.5 Sonnet (20240620) | 44.09 | #308 |
| 12 | GPT-5.4 | 43.42 | #83 |
| 13 | Grok 4 | 43.41 | #189 |
| 14 | Claude 3.7 Sonnet (20250219) | 42.17 | #209 |
| 15 | Claude 3.5 Haiku (20241022) | 42.16 | #590 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=humanagencybench · How It Works · Data refreshed daily, snapshot 2026-11-05.