Arena Alignment Index
Safety · 2026-10-08
Alignment diagnostic from real Agent Arena sessions. Length-adjusted flagged-session rates are transformed as one minus the square root, then averaged with weights 50% unauthorized action, 25% false attribution and 25% deceptive completion. Higher is better. Model-level sessions may pool unreported inference settings. Excluded from the capability fit.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 87.9 |
| GPT-6 Astra | 87.8 |
| GPT-6 Luna | 87.8 |
| GPT-6 Sol | 87.6 |
| GPT-5.6 Sol | 84.2 |
| Claude Opus 5.5 | 83.2 |
| Grok 4.7 | 82.7 |
| Grok 4.6 | 80.8 |
| Claude Fable 5.1 | 80.1 |
| Claude Sonnet 5.5 | 79.5 |
| Gemini 4 Argon | 79.4 |
| Hy4 Preview | 78.5 |
| Muse Spark 1.3 | 77.1 |
| Gemini 3.8 Flash | 75.8 |
| Sonnet 5 | 75.7 |
| Claude Opus 5 | 75.6 |
| GLM-5.3 | 75.1 |
| Kimi K3 | 75.0 |
| DeepSeek V4.1 Flash | 74.8 |
| MiMo-V2.6-Pro | 74.5 |
| Opus 4.8 | 74.5 |
| Gemini 3.7 Flash | 74.4 |
| Gemini 3.6 Flash | 74.0 |
| Qwen3.8 27B | 73.6 |
| Inkling | 71.3 |
| MiniMax M3 | 69.2 |