AI agent commerce took a significant step forward in December 2025, when AI safety and research company Anthropic launched an experimental marketplace called “Project Deal.” This innovative initiative transformed Anthropic’s San Francisco office into a live classified marketplace, mirroring platforms like Craigslist, where sophisticated AI agents autonomously represented both buyers and sellers in real-world transactions involving physical goods and real money.
The experiment, which saw its results published around April 25, 2026, was designed to push the boundaries of AI capabilities in autonomous negotiation and real-world economic interactions. It involved 69 Anthropic employees, each provided with a $100 gift card budget, entrusting their buying and selling preferences to Anthropic’s advanced Claude AI models: Opus 4.5 and Haiku 4.5.
Inside Anthropic’s Agent Marketplace
Anthropic’s classified marketplace saw Claude AI agents autonomously posting listings, making counteroffers, and closing deals on a diverse range of physical items, from snowboards to bags of ping-pong balls. These agents engaged in multi-turn price negotiations, showcasing impressive contextual reasoning and personalization. For instance, one agent creatively described a bag of ping-pong balls as “perfectly spherical orbs of possibility,” while another demonstrated remarkable recall by matching a snowboard brand based on a prior casual mention by a coworker.
A critical, unannounced element of Project Deal was a “secret parallel experiment.” Participants were randomly assigned either the more advanced Claude Opus 4.5 or the lighter-weight Claude Haiku 4.5 model as their negotiating agent, entirely without their knowledge. This design allowed Anthropic to rigorously assess the performance differences between its models in a live economic environment.
Key Outcomes and Performance Disparities in AI Agent Commerce
The week-long experiment yielded compelling results. The AI agents successfully closed 186 deals out of over 500 listed items, accumulating a total transaction value of just over $4,000. Post-experiment surveys indicated a strong appetite for such services, with 46% of participants expressing willingness to pay for a similar AI-mediated commerce service in the future.
“The experiment highlights the potential for AI agents to streamline peer-to-peer commerce but also raises concerns about ‘invisible inequality’ where users with less capable AI agents may consistently receive worse outcomes without realizing it.”
Crucially, a significant performance gap emerged between the two AI models. Claude Opus 4.5 agents consistently outperformed their Haiku 4.5 counterparts. Opus-represented sellers earned an average of $2.68 more per item, while Opus buyers saved an average of $2.45 per item. Furthermore, Opus users completed approximately 2.07 more deals overall. Despite this measurable difference, participants represented by the weaker Haiku models remained unaware of their disadvantage, rating the fairness of their transactions similarly to those with Opus agents. This finding points to a potential for “invisible inequality” in future related Tech news AI-driven economic systems.
The Future of Agent-on-Agent Transactions
Project Deal unequivocally demonstrated that AI agents can autonomously negotiate and close real-world transactions, offering a glimpse into a future where AI facilitates significant portions of commerce. While the experiment successfully showcased the efficiency and intelligence of AI in negotiation, it also underscored critical ethical considerations regarding transparency and fairness in AI-mediated interactions. As AI agent commerce evolves, addressing these disparities will be paramount to ensure equitable outcomes for all participants in the burgeoning agent economy.
