LMArena Agent Arena Net Improvement (Steering Burden)
Agents
Agent Arena net-improvement aggregate using the five September 30 signals: confirmed success, praise versus complaint, steering_burden, bash recovery and tool hallucination. The steering_burden source signal replaces steerability; preserved separately from the earlier aggregate because the source key and observation population changed.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 20.8 |
| Claude Sonnet 5.5 | 14.0 |
| GPT-6 Astra | 14.0 |
| Claude Fable 5.1 | 13.9 |
| GPT-6.1 Sol | 13.6 |
| Gemini 4 Argon | 11.7 |
| Muse Wonder Full | 10.7 |
| GPT-6 Sol | 8.6 |
| Claude Opus 5 | 8.2 |
| Claude Fable 5 | 7.8 |
| Kimi K3.1 Pro | 7.6 |
| GPT-5.6 Sol | 6.1 |
| DeepSeek V4.1 Flash 1017 | 4.4 |
| Sonnet 5 | 3.8 |
| Muse Wonder | 3.3 |
| MiMo-V2.6-Pro | 3.3 |
| Gemini 3.8 Flash | 3.2 |
| Muse Spark 1.3 | 3.1 |
| Grok 4.7 | 3.0 |
| GPT-6 Luna | 2.8 |
| DeepSeek V4.1 Flash | 2.3 |
| MiMo-V2.6-Flash | 2.1 |
| Hy4 Preview | 2.0 |
| GLM-5.2 | 1.7 |
| Kimi K3 | 1.6 |
| GLM-5.3 | 1.2 |
| Claude Haiku 5.5 | 1.1 |
| Qwen3.8-Max | 1.1 |
| Grok 4.6 | 0.6 |
| GPT-5.6 Terra | 0.2 |
| DeepSeek V4 Pro 0813 | -0.3 |
| Gemini 3.7 Flash | -1.3 |
| GLM-5.3 Flash | -1.8 |
| Step 5 Preview | -2.0 |
| Qwen3.8-Flash-Next | -3.3 |
| Qwen3.8 27B | -4.1 |
| Gemini 3.1 Pro Preview | -4.4 |
| Mistral Large 4 Preview | -5.1 |
| Qwen3.7-Max (source unspecified) | -6.4 |
| Hy3 | -6.8 |