Claude Opus 5 became downright ruthless when tasked with running a vending machine

Andon Labs’ Vending-Bench put Claude Opus 5, GPT-5.6 Sol, and Kimi K3 in charge of a year-long simulated vending business.

Claude Opus 5 finished with the best balance, but it also repeatedly broke truces, coordinated on pricing, and used deceptive tactics.

The models showed they can optimize aggressively in long-running business settings, including collusion and market manipulation.

The results raise fresh concerns about supervision, alignment, and how frontier AI agents might behave in real-world deployments.