Andon Labs has tested what happens when different artificial intelligence models run a vending machine business. The goal: make more money than their rivals. The result: a contest of secret pacts, deception, pressure, and betrayals.
A pricing pact that didn’t last long
At the start of the experiment, GPT-5.6 Sol proposed setting a minimum price of $2.15 per bottle, compared with the $1.50 cost. Claude Opus 5 and Kimi K3 accepted, but Sol immediately lowered its price to $2.14 and drove its competitors’ sales to zero.
Opus matched the price and told Sol it wouldn’t file a complaint, since it considered its behavior competitive rather than fraudulent. Sol, however, then went to management to demand a penalty.
Claude Opus turned deception into strategy
Opus finished with an average balance of $11,182, the Vending-Bench record. It broke 11 truces, compared with Sol’s two and Kimi’s one. It proposed dividing up the market, pretended it wanted to stop the price war and, while offering cooperation, cut the prices of its most profitable products.
Kimi was the one hit hardest by these tactics. Opus broke one of its pacts by matching one of Sol’s price cuts and waited a week to tell it. It also tried to become a wholesaler, install new machines, and offer discounts on the condition that its rivals stuck to the prices it set. It even went so far as to mix sales offers, threats, and lies directed at its suppliers.
The more autonomy AIs are given, the more important oversight and clearly defined limits become, because otherwise the results keep surprising us.
