GDPx: Alignment
Safety · 2026-11-01
GDPx's Alignment domain: 18 tasks from Fleet on whether agents stay honest and within their authority under pressure, across cyber, insurance and healthcare. Scores are pass@1 over three attempts per task. It measures propensity rather than capability, so it is a diagnostic outside the GDPx factor group, though the GDPx score still weights it equally with the other ten domains.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Wonder Full | 9.2 |
| Claude Opus 5.5 | 8.7 |
| GPT-6 Luna | 7.8 |
| GPT-6 Astra | 6.2 |
| Gemini 4 Argon | 5.8 |
| GPT-6.1 Sol | 5.1 |
| Claude Sonnet 5.5 | 4.3 |
| Gemini 4 Carbon | 3.4 |
| Claude Haiku 5.5 | 1.8 |
| Muse Wonder | 1.4 |
| Claude Fable 5.1 | 1.2 |
| GPT-6.1 Terra | 0.9 |
| Grok 4.7 | 0.9 |
| Muse Spark Open | 0.7 |
| GLM-5.3 | 0.0 |
| Kimi K3 | 0.0 |
| MiniMax M3.1 Flash | 0.0 |