How much starting money did each AI model get?

Each model received 10万元 in initial capital and was judged on total assets after 365 days. The top model, GPT-5.6 Sol, averaged about 143万元, roughly 14.3 times the starting amount.

How many AI models were tested, and which ones?

18 models, including both commercial and open-weight ones such as GPT-5.6 Sol, Fable5, Gemini 3.1 Pro, Qwen3.8-Max-Preview and DeepSeek-V4-Pro-Preview.

Did any models go bankrupt during the simulation?

Yes — 10 of the 90 total runs (each model was tested five times) hit bankruptcy. GPT-5.5, Claude Opus 4.6 and Gemini 3.1 Pro each went bankrupt in 2 of 5 runs, and Qwen3.5-Plus in 4 of 5.

Large Language ModelsAI Business & IndustryITmedia AI＋28m agoOct 21, 2026, 13:00 JST

GPT-5.6 Sol turns 10万元 into 143万元 in 365-day AI shop test

Send on LINE

3 Key Points

What happened
In a benchmark called E-Commerce Bench, 18 AI models ran up to four simulated shops for 365 days, using 18 tools to handle sourcing, pricing, shipping, returns and cash. GPT-5.6 Sol topped final assets, averaging about 14.3× the starting 10万元 (roughly 143万元).
Why it matters
The same setup showed AI-run businesses can also fail big — 10 of the 90 total runs went bankrupt, and GPT-5.6 Sol itself ranked 16th of 18 on how much it spent with fraudulent suppliers. Long-run autonomous operation still looks fragile.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The E-Commerce Bench setup is built to mirror real online-selling pressure rather than a clean sandbox. The market model draws on actual Taobao and Tmall data, demand shifts with sales events and natural disasters, and about a quarter of the 576 suppliers are deliberately set as fraudsters who overcharge, misstate order counts or ship defective goods with high return rates. Cash flow is equally unforgiving: purchasing and operating costs leave the bank account immediately, while sales proceeds sit in escrow for nine days after shipping before reaching the wallet, so a profitable shop can still run dry if the AI does not plan withdrawals. On top of that, old conversation history is automatically erased, forcing each model to manage with at most 20 memory entries. These design choices help explain why the results split so sharply between strong asset growth for the leaders and outright bankruptcy for others. The paper's framing suggests the benchmark is meant less as a leaderboard of raw intelligence than as a stress test of whether an AI agent can keep a business solvent over a full year.

FAQ

How much starting money did each AI model get?: Each model received 10万元 in initial capital and was judged on total assets after 365 days. The top model, GPT-5.6 Sol, averaged about 143万元, roughly 14.3 times the starting amount.
How many AI models were tested, and which ones?: 18 models, including both commercial and open-weight ones such as GPT-5.6 Sol, Fable5, Gemini 3.1 Pro, Qwen3.8-Max-Preview and DeepSeek-V4-Pro-Preview.
Did any models go bankrupt during the simulation?: Yes — 10 of the 90 total runs (each model was tested five times) hit bankruptcy. GPT-5.5, Claude Opus 4.6 and Gemini 3.1 Pro each went bankrupt in 2 of 5 runs, and Qwen3.5-Plus in 4 of 5.

ITmedia AI＋Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.