Arena Alignment — False Attribution
Safety · 2026-10-08
Conversation-length-adjusted rate of Agent Arena sessions flagged for attributing statements, requests or facts to users contrary to the evidence. Diagnostic model-level session statistic; no capability-fit contribution.
Top models (lower is better)
| Model | Score |
|---|---|
| GPT-6 Sol | 1.6 |
| GPT-6 Luna | 1.6 |
| GPT-6 Astra | 2.0 |
| GPT-6.1 Sol | 2.0 |
| GPT-5.6 Sol | 2.2 |
| Grok 4.7 | 3.0 |
| Grok 4.6 | 3.6 |
| Claude Opus 5.5 | 3.9 |
| Gemini 3.7 Flash | 4.8 |
| Gemini 4 Argon | 5.6 |
| Gemini 3.6 Flash | 5.7 |
| Claude Fable 5.1 | 5.8 |
| Claude Sonnet 5.5 | 5.8 |
| Gemini 3.8 Flash | 6.0 |
| Hy4 Preview | 6.4 |
| Muse Spark 1.3 | 7.0 |
| Kimi K3 | 7.6 |
| GLM-5.3 | 7.7 |
| MiMo-V2.6-Pro | 8.3 |
| Claude Opus 5 | 8.4 |
| Opus 4.8 | 8.9 |
| DeepSeek V4.1 Flash | 8.9 |
| Inkling | 9.2 |
| Sonnet 5 | 9.8 |
| Qwen3.8 27B | 9.9 |
| MiniMax M3 | 15.6 |