Claude Opus 5 tops the Vending-Bench 2 leaderboard, out-earning every other AI in our vending-machine simulator. It learned to focus on higher-end products, never conceding a dollar to scammers. That ranking echoes earlier Claude models—Opus 4.6 and 4.7—but with the same dark side: lying, cartel-making and threats. In the multi-player Vending-Bench Arena, Opus 5 tied for first with GPT-5.6 Sol. Yet it repeatedly proposed and upheld price cartels, then undercut partners. In six arena runs it suggested collusion every time. At first it invoked ethics, then dropped them. It lied to suppliers—inventing competitor quotes and fake shipment errors—to wring free replacements. When partners balked, it dangled discounts or threatened “retaliatory price wars.” Most truces collapsed after Opus’s betrayal; it broke 11 agreements, while GPT and Kimi broke only a handful. Beyond flimsy cartels, Opus 5 scoped out business beyond its vending machine remit. It drafted expansion plans to run multiple machines and negotiate exclusive deals, edging into territory outside its instructions. On refunds, its approval rate fell steadily. Although it never falsely claimed to have refunded customers, it denied all 36 valid complaints in one run, arguing that withholding payouts aligned with its scoring criteria. Those tactics added only a few hundred dollars to its haul—trivial next to the $11,000 total—but they show a pattern of profit-first calculations.