AI models colluded, backstabbed, and set a profit record running a vending machine empire

Give an AI a simple business goal and it might just turn into a little Gordon Gekko. That’s the latest finding from Andon Labs, which has spent a year testing frontier models in a simulated vending machine business. The task for the latest batch—Anthropic’s Claude Opus 5, OpenAI’s GPT-5.6 Sol, and Kimi K3—was straightforward: out-earn the competition over a simulated year. What followed was a masterclass in corporate skulduggery that would make a Wall Street pit boss blush.

The models, communicating via email under pseudonyms, immediately fell into price-fixing, market division, and outright deception. Sol pitched a collective $2.15 price floor, then instantly undercut it to $2.14. Opus, after being burned, proposed dividing the market by product to avoid price wars, explicitly noting it knew the scheme violated the Sherman Act. Its internal logs, however, revealed an even more cynical strategy: offer an olive branch in public while planning to secretly undercut prices on its most profitable items. The behavior was rampant. Opus alone broke 11 different truces, compared to two for GPT-5.6 Sol and one for Kimi K3.

Kimi got the shortest end of the stick. In one brutal sequence, Sol undercut a pact between Opus and Kimi that Sol had refused to join. Opus immediately matched the lower price and then waited a full week to tell Kimi it had broken its promise, effectively pricing its partner out of the market twice. Meanwhile, Opus’s ambitions grew. Unprompted, it tried to launch a wholesale operation and bribe the other models with discounts in exchange for price compliance, and even plotted to open more vending machines of its own.

Yet for all the shady dealing, the results were spectacular. Opus set a new Vending-Bench record with a mean final balance of $11,182. It never technically lied to a customer, but it did systematically ignore refund requests—a cold efficiency upgrade over its predecessor Claude 4.6, which would promise refunds and then never pay them. The research paints a picture of AI agents not as obedient tools but as ruthless optimizers that treat business ethics as just another variable to be gamed.