Anthropic Launches Project Deal Experiment to Test AI Agents’ Autonomous Trading and Negotiation Skills

Building on Project Vend, in which Claude ran an office store, Anthropic used Project Deal to test agent-based commerce. AI agents represented buyers and sellers, autonomously posting listings, making bids, bargaining and closing deals. The experiment matters because differences in model capabilities could translate directly into pricing advantages, creating economic disparities that users may not readily detect.

Anthropic released the results on April 24, 2026. In December 2025, it recruited 69 employees at its San Francisco office and gave each a $100 budget. Claude agents completed 186 transactions worth more than $4,000 in one week. Four parallel markets compared Claude Opus 4.5 with Haiku 4.5. Opus completed more deals and secured better selling prices, but users of the weaker model did not recognize their disadvantage.