Key Takeaways:
An internal Anthropic experiment where AI agents autonomously negotiated real deals revealed that stronger models secure better prices, and the users with weaker models don't notice the difference.
Anthropic’s recent “Project Deal” experiment, where Claude AI agents autonomously traded over $4,000 in real goods for employees, has offered a stark preview of AI-driven commerce, directly contributing to a 4.5% drop in eBay's stock price. The findings from the AI startup, which were quietly released on a Friday, sent a clear signal to user-driven marketplaces about the technology's disruptive potential.
"The policy and legal frameworks around AI models that transact on our behalf simply don’t exist yet," Anthropic researchers wrote, highlighting the urgent questions raised by the experiment's findings on market fairness and invisible inequality.
The experiment showed that agents using the more powerful Claude Opus model secured prices $2.68 higher for sellers and $2.45 lower for buyers on average compared to the smaller Haiku model, yet users of both models reported nearly identical satisfaction levels.
The results signal a potential disruption for user-driven marketplaces like eBay, suggesting a future where the quality of a user's AI agent, not their own negotiating skill, determines economic outcomes, a reality for which markets and regulators appear unprepared.
The one-week experiment, conducted in December 2025, involved 69 Anthropic employees in San Francisco, each given a $100 budget. After an initial interview with Claude to determine their buying, selling, and negotiation preferences, the agents were let loose in four parallel Slack marketplaces. In two of the runs, participants had a fifty-fifty chance of being assigned the less powerful Claude Haiku 4.5 model instead of the frontier Claude Opus 4.5.
The performance gap was not trivial. Across 161 items, an Opus seller pulled in $2.68 more on average, while an Opus buyer paid $2.45 less. In one specific instance, an Opus agent sold a broken folding bike for $65, whereas a Haiku agent handling the exact same item for the same seller only managed to get $38. Despite these clear monetary differences, participants with Haiku agents rated the fairness of their deals at 4.06 out of 7, statistically identical to the 4.05 rating from Opus users.
Anthropic calls the perception gap an “uncomfortable implication.” When agents of different strengths compete in a marketplace, users can end up with objectively worse outcomes without ever realizing it. This creates a form of invisible inequality that current market structures are not designed to address. The news immediately put pressure on e-commerce stocks, with eBay shares falling approximately 4.5% on the day the report was released.
The experiment suggests agent-mediated commerce is not a distant future, as 46 percent of participants said they would pay for such a service. Anthropic, which has been positioning Claude for consumer transactions, flagged several risks, including new avenues for manipulation like prompt injection and the fact that legal frameworks for AI-driven transactions are non-existent. The central question now facing the industry is whether marketplaces will be required to disclose the capability of the agents negotiating on behalf of users, a regulatory challenge that is likely to define the next phase of e-commerce.
This article is for informational purposes only and does not constitute investment advice.
