LMArena Agent Arena Net Improvement
Agents · 2026-06-04
Net-improvement score for Agent Arena, summarizing the model’s effect on session outcomes relative to a baseline. Higher net improvement indicates better aggregate agentic value in live arena sessions.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 12.6 |
| Claude Opus 5 | 11.9 |
| GPT-5.6 Sol | 10.0 |
| Kimi K3 | 9.9 |
| Opus 4.8 | 9.4 |
| Sonnet 5 | 8.6 |
| GPT-5.5 | 8.5 |
| Opus 4.7 | 8.2 |
| GLM-5.2 | 7.1 |
| Opus 4.6 | 6.5 |
| Grok 4.5 | 5.5 |
| GPT-5.4 | 5.1 |
| GPT-5.6 Terra | 3.5 |
| GPT-5.6 Luna | 3.5 |
| Sonnet 4.6 | 3.2 |
| Muse Spark 1.1 | 0.7 |
| GLM-5.1 | 0.4 |
| Kimi K2.7 Code | 0.4 |
| Gemini 3.1 Pro Preview | -0.5 |
| Qwen3.7-Max | -0.5 |
| DeepSeek-V4-Pro | -0.8 |
| Hy3 | -1.0 |
| Gemini 3.5 Flash | -1.0 |
| Kimi K2.6 | -1.1 |
| Gemini 3.6 Flash | -1.6 |
| Qwen3.7-Plus | -1.7 |
| MiMo-V2.5-Pro | -2.5 |
| MiniMax M3 | -2.6 |
| DeepSeek-V4-Flash | -3.0 |
| Inkling | -5.7 |
| Grok Build 0.1 | -8.0 |
| Grok 4.3 | -8.0 |
| Gemini 3 Flash Preview | -8.8 |
| Gemini 3.5 Flash-Lite | -10.2 |
| MiniMax M2.7 | -11.6 |
| Nemotron 3 Ultra 550B A55B | -13.1 |
| Gemma 4 31B IT | -16.1 |