Top AI Model Grows Simulated Online Shop Capital 14.3-Fold in a Year

100,000 yuan, about 2.35 million yen. The AI is given this sum and told to run an online shop for a year. A paper titled “E-Commerce Bench,” compiled by researchers from Alibaba's Qwen team and the Hong Kong University of Science and Technology (HKUST), presents a benchmark (a common test that compares multiple models under the same conditions) that measures the autonomous business capabilities of AI models through this kind of long-term simulated operation. The paper is a preprint posted on arXiv.

The setting is a market model based on data from real e-commerce platforms, Taobao and Tmall. The market contains 6,886 products and 576 suppliers, and the AI operates up to four stores at the same time. Using 18 types of tools, it autonomously handles everything from market research to price negotiation, procurement, pricing, shipping, returns processing and cash management. About one-quarter of the suppliers are set up as scammers that overcharge relative to market prices, falsify order numbers or ship low-quality goods with soaring return rates.

The cash flow also has traps. Procurement and operating costs are debited immediately from the bank account, while sales proceeds are held for nine days after shipment in an escrow account (a system in which a third party holds payment to secure the transaction) before they reach the wallet. Therefore, even a profitable store will go bankrupt if it does not plan withdrawals to its bank account, as the bank balance remains negative for 10 consecutive days. In addition, old conversation history is automatically deleted once it exceeds a certain amount, so the AI must use up to 20 memory entries to keep track of transaction information.

The top performer was “GPT-5.6 Sol,” which grew its initial funds by about 14.3 times (about 1.43 million yuan) on average. However, it ranked 16th of 18 models for the share of spending directed to scam suppliers. “Fable5” (about eight times), in second place, earned steadily with less than half the number of tool calls used by GPT-5.6 Sol, and ranked first in profit efficiency per action. “Claude Opus 4.7” showed the most meticulous performance among all models in both price negotiation and scam avoidance, but it failed to expand its trading volume, leaving its total assets in eighth place. The best open-weight model was Alibaba's “Qwen3.8-Max-Preview” (about 4.2 times, fifth overall), which had the top long-term learning capacity among all 18 models, using the lowest past purchase price as a benchmark to win discounts when restocking.

These results matter to businesses that want to hand over operations to AI and to developers choosing models. The paper shows with numbers that the models that grow assets the most are not necessarily the ones that avoid scams and bankruptcy. Of the 90 trials in total, 10 ended in bankruptcy partway through. By model, GPT-5.5, Claude Opus 4.6 and Gemini 3.1 Pro went bankrupt in two of five runs each, and Qwen3.5-Plus in four of five. However, each model was tested only five times, and the market is simulated.