Some AI models increased their initial assets more than 14-fold in an experiment that put them in charge of operating an online shopping mall. Other models went bankrupt after failing to manage funds, confirming that AI's ability to run a business autonomously varies widely by model.

On Tuesday, Japanese media outlet ITmedia reported that Alibaba's Qwen research team and a joint research team at the Hong Kong University of Science and Technology announced results from E-Commerce Bench, a study evaluating AI's long-term management capabilities.

Researchers provided each of 18 AI models with initial funds of 100,000 yuan (19,844,000 won) and had them operate a virtual online shopping mall for 365 days. It was a simulation using e-commerce data, not an actual year-long business operation.

The AIs ran up to 4 stores and handled market research, supplier negotiations, product procurement, pricing, shipping, returns processing and fund management. The test included 6,886 products and 576 suppliers.

In particular, 152 suppliers were set up as fraudulent firms that lied about prices or delivered defective goods. Sales proceeds were also locked up for 9 days after delivery, meaning that even profitable operations could go bankrupt if they failed to manage cash flow.

OpenAI's GPT-5.6 SOL ranked first by final assets. It increased the initial 100,000 yuan to an average of about 1.43 million yuan (283.77 million won), a 14.3-fold gain. Its share of spending on fraudulent firms was weak, ranking 16th among the 18 models.

Anthropic's Fable5 ranked second, with final assets of about 800,000 yuan (158.73 million won). It recorded high profit efficiency with fewer tool calls than GPT-5.6 SOL.

Among open-weight models, Alibaba's Qwen3.8 Max Preview delivered the highest performance at about 420,000 yuan (83.33 million won). It also stood out for its ability to remember past transaction prices and use them in renegotiations.

Of 90 experiments overall, 10 ended in mid-course bankruptcy. Some models halted operations due to insufficient funds in 4 out of 5 attempts.

The researchers said the test showed the possibility of AI running businesses over long periods, but limits remained in areas such as fraud detection and fund management.