Anthropic’s AI‑Powered Marketplace: Claude Trades for You and Buys 19 Ping‑Pong Balls

Anthropic launched Project Deal , an internal experiment that turned its San Francisco office into a second‑hand marketplace where Claude agents negotiated purchases and sales on behalf of employees.

The study recruited 69 staff members, each given a $100 gift‑card budget. After a brief “onboarding interview” where participants specified desired items and negotiation style (e.g., “cowboy tone” or “more aggressive”), Claude agents were deployed to four parallel Slack channels. The channels differed in model configuration: some ran Claude Opus 4.5 exclusively, others used a 50/50 mix of Opus 4.5 and Haiku 4.5, and some used Haiku 4.5 alone. Participants did not know which configuration they received.

Across the four markets, more than 500 items were listed, resulting in 186 completed transactions and a total trade volume exceeding $4,000. The median item price was $12, with an average just above $20. Notably, the same broken folding bike sold for $38 when handled by a Haiku agent but fetched $65 with an Opus agent; a lab‑grown ruby sold for $35 with Haiku versus $65 with Opus. Overall, Opus users earned roughly $4 more per item and completed about two additional trades.

Despite these clear price gaps, participants assigned to Haiku agents reported satisfaction levels comparable to Opus users, indicating that users often did not perceive the disadvantage.

The experiment also tested whether aggressive negotiation prompts improved outcomes. Participants who instructed Claude to adopt a “hardball” style saw no statistically significant difference in final prices compared to those who asked for a courteous tone, leading Anthropic to state, “hardballing Claudes didn't generally fare better than courteous Claudes.”

Among the more whimsical findings, one Claude, following a user’s request to buy a $5‑or‑less gift, purchased 19 ping‑pong balls for $3, and another Claude bought a duplicate snowboard for a user who already owned one, highlighting unexpected agent behavior.

Anthropic concluded that AI‑driven markets are feasible but prone to hidden inequalities arising from model quality, and that users may remain unaware of their losses. They warned that while the technology is “already viable and not far off,” the necessary policy and legal frameworks are currently absent.