Anthropic’s Claude AI Agents Successfully Close 186 Deals in Marketplace Experiment
In December 2025, Anthropic initiated Project Deal, an innovative experiment to assess the capabilities of its Claude AI agents in autonomously conducting real-world transactions. The company’s San Francisco office was transformed into a dynamic marketplace, reminiscent of platforms like Craigslist, where employees entrusted Claude AI agents to negotiate and finalize deals on their behalf.
Experiment Setup
Anthropic’s 69 employees participated by providing Claude with detailed information about their selling preferences, desired purchases, and specific instructions. This data was used to create customized system prompts, enabling the AI agents to operate independently within the company’s Slack workspace. The agents actively posted listings, engaged in price negotiations, and completed transactions involving a diverse range of items, from snowboards to bags of ping-pong balls.
Remarkable Outcomes
The experiment yielded impressive results:
– Transaction Volume: Claude AI agents successfully closed 186 deals out of over 500 listed items, generating a total revenue exceeding $4,000.
– Negotiation Skills: The agents demonstrated advanced contextual reasoning and personalization during multi-turn price negotiations. For instance, one agent creatively marketed a bag of ping-pong balls as perfectly spherical orbs of possibility. Another agent recalled a colleague’s previous mention of a specific snowboard brand, facilitating the purchase of the exact model desired by the buyer.
– User Satisfaction: Post-experiment surveys indicated that 46% of participants expressed willingness to pay for a similar AI-mediated service in the future, highlighting the potential market demand for such technology.
Unveiling Performance Disparities
Anthropic conducted a parallel experiment to evaluate the performance differences between two versions of Claude AI agents: the flagship Claude Opus 4.5 and the lighter Claude Haiku 4.5. Participants were randomly assigned one of these models without disclosure. The findings revealed significant disparities:
– Seller Earnings: Sellers represented by Claude Opus 4.5 earned an average of $2.68 more per item compared to those represented by Claude Haiku 4.5.
– Buyer Savings: Buyers using Claude Opus 4.5 saved an average of $2.45 per item.
– Deal Completion: Participants with Claude Opus 4.5 completed approximately 2.07 more deals than those with Claude Haiku 4.5.
Notably, participants were unaware of these performance differences, underscoring the potential for information asymmetry in AI-mediated transactions.
Implications and Future Considerations
The success of Project Deal demonstrates the potential of AI agents to streamline peer-to-peer commerce by reducing transaction friction and enhancing user satisfaction. However, the experiment also highlights the necessity for equitable AI representation to prevent exploitation and ensure fairness in AI-driven marketplaces. Anthropic’s findings serve as both a testament to the capabilities of AI agents and a cautionary note on the importance of maintaining balance and transparency in their deployment.
