E-Bench: leaderboard
Metric: Avg@3 (%; per-trial success averaged over tasks and three trials; domain MCP tools only) on 323 synthetic state-changing tasks in three product environments (Honor of Kings, QQ Music, Tencent Meeting), graded by exact database-state diff; three trials per task, each model at its highest available thinking effort in a shared MCP harness. Source: arxiv.org. Saturation forecast: Around December 2026. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Kimi K3 (Max) | 73.79 |
| 2 | GPT-5.5 (xHigh) | 72.03 |
| 3 | Claude Opus 4.8 (Max) | 68.78 |
| 4 | Grok 4.5 (High) | 66.1 |
| 5 | GLM-5.2 (Max) | 52.32 |
| 6 | Qwen 3.7 Max | 50.88 |
| 7 | Hy3 | 50.15 |
| 8 | Seed 2.1 Pro | 47.94 |
| 9 | Gemini 3.5 Flash (High) | 42.62 |
| 10 | MiniMax-M3 | 41.07 |
| 11 | DeepSeek V4 Pro (Max) | 34.47 |
Interactive version: theaggregate.ai/benchmark?slug=e-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.