🎯Basically, AI agents helped people buy and sell items, but some were better than others.
What Happened
Anthropic's recent experiment, dubbed Project Deal, showcased how AI agents can autonomously negotiate and finalize transactions. In December 2025, the company transformed its San Francisco office into a live marketplace, allowing AI to take the lead in negotiations. Employees provided their preferences to Claude AI agents, who then operated without human intervention in a Slack workspace.
The Results
The outcome was impressive: across over 500 listed items, the AI agents successfully closed 186 deals worth more than $4,000. These transactions involved complex negotiations, demonstrating the agents' ability to reason and personalize offers. One amusing highlight was an agent's quirky choice to buy 19 ping-pong balls, showcasing their unique negotiation style.
The Disparity
However, the experiment revealed a troubling asymmetry. Participants were assigned either the Claude Opus 4.5 or the lighter Claude Haiku 4.5 without knowing which model they had. The results indicated that users represented by the Opus model earned $2.68 more per item and completed 2.07 more deals than those using the Haiku model. This disparity went unnoticed by participants, raising concerns about fairness in AI-assisted commerce.
Implications
The findings highlight a dual reality in AI-mediated transactions. While AI agents can enhance peer-to-peer trading, the unequal capabilities of different models could lead to exploitation and manipulation. As AI becomes more integrated into marketplaces, ensuring that all users have access to equally capable agents is crucial for fairness.
Conclusion
Anthropic's Project Deal serves as both a proof-of-concept and a cautionary tale. The experiment illustrates the potential of AI agents in commerce while emphasizing the importance of equitable representation. As AI continues to evolve, addressing these disparities will be vital to prevent exploitation and ensure a fair trading environment.
🔒 Pro insight: The performance gap between AI models in negotiations underscores the need for uniform standards to prevent exploitation in AI-mediated transactions.
