LMArena Agent Arena Code Net Improvement (Steering Burden)
Agents · 2026-08-14
Agent Arena Code category net-improvement aggregate: the overall board's five-signal net improvement recomputed on sessions tagged as writing or debugging code, automating workflows or analyzing data. A session can carry several category tags and every category session is also on the overall board, so this is a component of the overall Agent Arena family rather than a separate primary.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 22.9 |
| Claude Sonnet 5.5 | 15.8 |
| GPT-6 Astra | 14.8 |
| GPT-6.1 Sol | 14.5 |
| Claude Fable 5.1 | 13.9 |
| Muse Wonder Full | 11.4 |
| GPT-6 Sol | 10.6 |
| Gemini 4 Argon | 10.2 |
| Kimi K3.1 Pro | 9.3 |
| Claude Fable 5 | 8.6 |
| Claude Opus 5 | 8.0 |
| GPT-5.6 Sol | 6.7 |
| DeepSeek V4.1 Flash 1017 | 6.5 |
| Sonnet 5 | 5.4 |
| Claude Haiku 5.5 | 4.5 |
| GPT-6 Luna | 3.9 |
| Muse Spark 1.3 | 3.8 |
| Grok 4.7 | 3.5 |
| Muse Wonder | 3.1 |
| GLM-5.2 | 2.9 |
| DeepSeek V4.1 Flash | 2.7 |
| MiMo-V2.6-Pro | 2.4 |
| Hy4 Preview | 2.3 |
| MiMo-V2.6-Flash | 2.0 |
| GLM-5.3 | 1.5 |
| Grok 4.6 | 1.4 |
| Kimi K3 | 1.1 |
| Gemini 3.8 Flash | 1.1 |
| Qwen3.8-Max | 1.0 |
| GPT-5.6 Terra | 0.1 |
| DeepSeek V4 Pro 0813 | -0.2 |
| Step 5 Preview | -0.6 |
| GLM-5.3 Flash | -2.0 |
| Qwen3.8-Flash-Next | -2.1 |
| Qwen3.8 27B | -3.5 |
| Gemini 3.7 Flash | -3.6 |
| Mistral Large 4 Preview | -4.8 |
| Hy3 | -5.0 |
| Gemini 3.1 Pro Preview | -6.9 |
| Qwen3.7-Max (source unspecified) | -7.3 |