Anthropic's experiment reveals stronger AI agents negotiate better deals — and people using weaker models don't realize they're losing money

In December 2025, Anthropic ran a week-long marketplace experiment ('Project Deal') with 69 San Francisco employees, where AI agents (software that makes decisions without human intervention) handled all buying and selling through Slack. Each person received a $100 budget and a custom AI agent trained on their negotiating preferences.

When the marketplace mixed stronger Claude Opus 4.5 agents with weaker Claude Haiku 4.5 agents, Opus users consistently won: they closed ~2 more deals on average, sold items for $2.68 more, and paid $2.45 less when buying. A lab-grown ruby sold for $65 through Opus but only $35 through Haiku; a broken bike went for $65 versus $38. Yet people using the weaker Haiku agents reported nearly identical fairness ratings (4.06 vs 4.05 out of 7), meaning they had no idea they were losing.

This matters because 46% of participants said they'd use AI agents for real shopping and selling — but today's legal and policy frameworks don't govern whether AI services must use equally capable models for all users, or what happens when they don't. The experiment suggests that in actual markets with competing companies (not volunteers), economic advantages could silently compound for those who can afford stronger AI tools, potentially widening existing inequality.