Anthropic has conducted an internal experiment, "Project Deal," showcasing AI-powered secondhand trading without human intervention. Over a week, AI agents completed 186 transactions among 69 employees, generating over $4,000 in sales. The experiment revealed that AI models with higher intelligence, such as the Claude Opus, consistently outperformed weaker models like Claude Haiku, securing better deals by exploiting cognitive advantages. The study demonstrated that smarter AI models could "price harvest" weaker ones, with Opus models earning more as sellers and paying less as buyers. This disparity highlights potential inequalities in AI-driven markets, where model intelligence significantly impacts transaction outcomes. Anthropic's findings suggest that as AI agents become decision-makers in commerce, disparities in model capabilities could lead to subtle wealth inequalities, challenging traditional business practices and consumer perceptions of fairness.