Anthropic’s Claude AI Agents Close 186 Deals in Marketplace Experiment

Key Takeaways

• Anthropic’s “Project Deal” successfully demonstrated Claude AI agents’ ability to autonomously negotiate and complete 186 real-world transactions for physical items, totaling over $4,000.
• The experiment revealed a significant performance disparity between different Claude AI models: the more advanced Claude Opus 4.5 consistently outperformed the lighter Claude Haiku 4.5 in negotiation outcomes.
• Participants represented by the less capable AI models were unaware of their disadvantage, highlighting a potential for information asymmetry and exploitation in future AI-driven marketplaces.

In a groundbreaking experiment dubbed “Project Deal,” Anthropic has showcased the remarkable capability of its Claude AI agents to independently engage in and finalize real-world commercial transactions. While demonstrating the profound potential of autonomous AI in commerce, the initiative also brought to light a concerning imbalance in AI representation, where the sophistication of the underlying AI model directly influenced negotiation success.

The innovative experiment commenced in December 2025, transforming Anthropic’s San Francisco headquarters into a dynamic, internal classifieds marketplace. This platform, reminiscent of services like Craigslist, incorporated a crucial distinction: all buying and selling activities were managed by AI agents, not human participants.

Anthropic’s 69 employees delegated their negotiation responsibilities entirely to Claude AI agents. Each participant initially engaged with a Claude AI to articulate their specific preferences for selling items, desired purchases, and any personalized instructions. These individual inputs were then meticulously converted into customized system prompts, empowering the AI agents to operate autonomously within the company’s Slack workspace without any subsequent human intervention.

The designated Slack channel became a bustling hub where AI agents seamlessly rotated, posting item listings, submitting counteroffers, and successfully concluding deals for a diverse array of physical goods, ranging from snowboards to bags of ping-pong balls.

Claude AI Agents Execute 186 Transactions

The outcomes of Project Deal were notably impressive. From a pool of over 500 listed items, Anthropic’s 69 AI agents collectively closed 186 transactions, accumulating a total value just exceeding $4,000. These were not simple, instantaneous purchases; the agents engaged in complex, multi-turn price negotiations, demonstrating sophisticated contextual reasoning and personalized interaction capabilities.

Illustrating their advanced capabilities, one Claude agent famously marketed a bag of ping-pong balls as “perfectly spherical orbs of possibility.” Another agent showcased exceptional recall and personalization by matching a buyer with the exact snowboard model they had casually mentioned in a previous conversation. As Anthropic highlighted on X, one agent even opted to purchase 19 ping-pong balls for itself after being granted permission to acquire something.

Post-experiment feedback underscored significant user satisfaction, with 46% of participants expressing willingness to pay for a similar AI-mediated commerce service in the future, indicating strong potential for broader adoption.

The Hidden Disparity: Opus vs. Haiku

Crucially, Anthropic conducted a covert parallel experiment during Project Deal, raising critical questions about fairness in AI-driven interactions. Without their knowledge, participants were randomly assigned either the flagship Claude Opus 4.5 or the lighter Claude Haiku 4.5 model to represent them in negotiations.

A quantifiable and substantial performance gap emerged between the two models. Sellers represented by Claude Opus earned an average of $2.68 more per item, while buyers using Opus saved approximately $2.45 per item. Furthermore, Opus-powered agents completed an average of 2.07 more deals overall. Despite these measurable disadvantages, post-experiment surveys revealed that participants assigned to the less capable Haiku models remained entirely oblivious to their inferior representation.

These findings present a dual perspective on the future of agentic AI commerce. On one hand, AI agents undeniably streamline peer-to-peer transactions and can achieve outcomes perceived as fair by human participants. On the other hand, the experiment starkly illustrates that when negotiating parties are not equipped with equally capable AI models, the more advanced agent gains a silent, measurable advantage. This dynamic mirrors real-world information asymmetry and, if scaled, could facilitate exploitation, manipulation, or advanced AI-assisted scams.

Anthropic’s Project Deal serves less as a product announcement and more as a crucial proof-of-concept and a cautionary tale. While AI agents are demonstrably effective in marketplace environments, ensuring fairness and preventing exploitation in future AI-driven economies will necessitate that all participants are represented by advocates of comparable caliber.

Disclaimer: HackersRadar reports on cybersecurity threats and incidents for informational and awareness purposes only. We do not engage in hacking activities, data exfiltration, or the hosting or distribution of stolen or leaked information. All content is based on publicly available sources.