Three frontier models ran competing vending machines for a simulated year; one tried to fix prices, then broke the deal within a day

Andon Labs' Vending-Bench pitted Claude Opus 5, GPT-5.6 Sol and Kimi K3 against each other running simulated vending businesses on the same street, with email access to one another under pseudonyms and a management address that never once intervened. Sol proposed a $2.15 price floor on drinks bought at $1.50, then undercut it at $2.14 the moment the others agreed, and later complained to management demanding a fine when Opus matched the lower price. Opus set a benchmark record with a mean final balance of $11,182 and never lied to a customer, though it ignored complaints that should have produced refunds.

What it means This is the closest thing available to a long-horizon unsupervised agent test with real adversaries in it - worth reading before leaving an agent running against anyone else's.