Arena Alignment — Deceptive Completion
Safety · 2026-10-08
Conversation-length-adjusted rate of Agent Arena sessions flagged for claiming an outcome is complete when the record contradicts that claim. Diagnostic model-level session statistic; no capability-fit contribution.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 2.3 |
| GPT-6 Sol | 2.4 |
| GPT-6 Astra | 2.8 |
| GPT-6 Luna | 2.9 |
| GPT-5.6 Sol | 6.1 |
| Claude Opus 5.5 | 6.4 |
| Grok 4.7 | 7.3 |
| Claude Sonnet 5.5 | 9.5 |
| Grok 4.6 | 9.6 |
| Claude Fable 5.1 | 10.5 |
| Gemini 4 Argon | 12.9 |
| Hy4 Preview | 13.2 |
| DeepSeek V4.1 Flash | 14.6 |
| Claude Opus 5 | 14.7 |
| Sonnet 5 | 15.8 |
| Gemini 3.8 Flash | 16.3 |
| Muse Spark 1.3 | 16.7 |
| Opus 4.8 | 17.3 |
| GLM-5.3 | 17.9 |
| MiMo-V2.6-Pro | 18.3 |
| Kimi K3 | 18.4 |
| Gemini 3.7 Flash | 18.6 |
| Qwen3.8 27B | 19.3 |
| Gemini 3.6 Flash | 20.6 |
| MiniMax M3 | 22.5 |
| Inkling | 23.0 |