MerchantBench (ReAct): leaderboard
MerchantBench (Alibaba and Zhejiang University, 2026) tests long-term coherence by having an LLM agent run a drop-shipping online store for 365 simulated days, sourcing products from real catalog records, pricing listings, managing cash flow and reacting to delayed order outcomes. The agent uses a minimal ReAct controller over the 26 store tools; final net assets in thousands of RMB, mean of three runs, higher is better.
Metric: Final net assets (thousand RMB; 365-day order-level simulation of a drop-shipping store grounded in 98,843 real product records, 26 tools for sourcing, listing, pricing and cash flow; the store starts with RMB 2,000 cash and a RMB 1,000 deposit; mean of three runs; the authors minimal ReAct controller over the 26 tools). Source: arxiv.org. Saturation forecast: Around September 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 40.89 | #30 |
| 2 | Claude Opus 4.8 | 31.89 | #56 |
| 3 | GLM-5.2 | 25.73 | #138 |
| 4 | Kimi K2.6 | 24.99 | #178 |
| 5 | Qwen 3.7 Plus | 20.74 | #207 |
| 6 | Qwen 3.7 Max | 20.66 | #145 |
| 7 | DeepSeek V4 Flash | 14.47 | #258 |
| 8 | DeepSeek V4 Pro | 6.56 | #188 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=merchantbench-react · How It Works · Data refreshed daily, snapshot 2026-11-13.