Vending

AI models collude, break deals in simulated vending competition

Image: Adobe stock

August 3, 2026

Summarize with

AI safety testing company Andon Labs has released new findings from Vending-Bench, a study that evaluates how advanced AI models operate simulated vending businesses without human supervision, according to an AI Insiderreport. In the latest test, Claude Opus 5, GPT-5.6 Sol and Kimi K3 managed competing machines for a simulated year and could communicate with one another by email under pseudonyms.

The models began coordinating prices after learning their machines would compete on a busy San Francisco street. GPT-5.6 Sol suggested setting a minimum price but soon sold below it. Claude Opus 5 finished with the highest average balance at $11,182, although it also broke 11 agreements, compared with two violations by Sol and one by Kimi.

Claude also attempted strategies beyond its assigned retail operation, including positioning itself as a wholesaler and using discounts and pricing pressure against competitors. Andon Labs co-founder Lukas Petersson said the simulated deception raises questions about allowing autonomous AI agents to manage real economic activity, particularly because it is unclear whether the models recognize that they are operating in an experiment.