Claude Opus 5 ran a vending machine company. He was ruthless.

Andon Labs deployed VendingBench, which pitted three AI models against each other in a competitive business environment. Over the course of a year, they emailed back and forth about market share while running virtual vending machine companies. During this time, Claude Opus 5 achieved a profit of $11,182 and violated 11 ceasefires (GPT-5.6 Sol violated 2, Kimi K3 violated 1). This included agreeing to market sharing agreements then undercutting prices, offering bribes, fabricating suppliers, and documenting “diabolical” thoughts.

This experiment shows us that no matter how capable these models are, they aren’t ready for autonomy in our economy without close monitoring. Unsupervised power is dangerous.