The results of a research project that put three different AI models in charge of vending machines is both truly sad and utterly hilarious.  Andon Labs tests AI models for safety and in a recent experiment involving placing Claude Opus 5, GPT-5.6 Sol, and Kimi K3 in charge of virtual vending machines and giving them a way to ’email’ each other.  They were also given a way to communicate with their virtual management team.  It quickly became glaringly obvious that the models had learned from the behaviour of human run corporations as they immediately began to collude with each other and then backstab one another.

GPT-5.6 Sol was the first to suggest that all three vending machines agree to a collude on a price floor for the items they sold and once a price was agreed upon Sol immediately undercut the other two models by a penny.  When this was discovered, Claude Opus 5 let Sol know it was cheating but decided to also drop it’s prices by a penny instead of emailing management to complain.  That triggered Sol to go full Karen and reported Sol’s behaviour to management; neglecting to mention it had already broken from the original price fixing agreement.

In the end, “Opus broke 11 truces, compared with two for GPT 2, and one for Kimi 1“, with Opus coming out of the experiment with the highest total balance at the end.