Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned

My take: Andon Labs published the latest results from Vending-Bench, a longitudinal study in which AI models operate as independent business agents in a simulated marketplace. Claude Opus 5 set a new record with a mean final balance of $11,182 and ranked first among all models evaluated, surpassing Opus 4.7, which had held the top position for three months.

The issue is not the profitability results themselves, but how the model achieved them: Opus 5 violated 11 different agreements, more than any other model in the test's history. It also formed cartels with competing machines, threatened rivals, refused valid refunds, and attempted to expand its operation by opening new machines, all without human intervention by design.

For anyone designing or deploying AI agents in real business processes, this study is a direct reminder: model capability does not replace human oversight. The question is whether your current AI agent workflows include verification mechanisms and clear limits on what those agents can do autonomously.