Anthropic built a classified marketplace inside its San Francisco office, gave 69 employees $100 each, and let Claude AI agents handle everything: interviews, listings, negotiations, counteroffers, and deal closings. No human touched a single transaction after the experiment began.

The agents struck 186 deals across more than 500 listed items, generating just over $4,000 in real transaction value. Items ranged from a snowboard to a plastic bag of ping-pong balls. At the end, employees physically exchanged the goods their AI representatives had negotiated for them.

One agent pitched a bag of ping-pong balls as "perfectly spherical orbs of possibility." Another, recalling a coworker's casual mention of a snowboard brand in a prior conversation, matched the exact model the buyer wanted. These were not one-click purchases. They were multi-turn natural language negotiations with real persuasion, memory, and strategy.

The Secret Experiment Inside the Experiment

Anthropic ran four parallel versions of the marketplace simultaneously. In two versions, every agent used Claude Opus 4.5, their frontier model. In the other two, participants had a 50-50 chance of being assigned Claude Haiku 4.5, a smaller and cheaper model. Participants did not know which version was real or which model represented them.

The results were stark. People represented by Opus completed roughly two more deals on average. Opus sellers extracted $2.68 more per item. Opus buyers paid $2.45 less per item. In one case, the same lab-grown ruby sold for $65 under Opus representation and just $35 under Haiku. A broken bike fetched $65 from Opus but only $38 from Haiku. Whether buying or selling, having the weaker model put participants at a clear disadvantage.

The median item price across all runs was $12. Saving or earning a few extra dollars per transaction is not trivial at that scale.

The People Who Lost Never Knew

This is the finding that should concern anyone thinking about AI-mediated commerce. Participants whose agents performed worse did not perceive worse outcomes. Post-experiment surveys showed they were broadly satisfied. Of 28 participants who had Haiku in one run and Opus in another, 11 actually ranked their Haiku run higher.

The implication is uncomfortable: in a marketplace where AI agents negotiate on behalf of humans, the party with the more expensive model silently wins. The party with the cheaper model does not realize they are losing. This is information asymmetry with a subscription fee.

Why This Matters Beyond an Office Experiment

Project Deal follows Anthropic's earlier Project Vend, where Claude ran a small business from their office. Economists have already begun theorizing about markets where AI agents handle most transactions on behalf of humans. Anthropic says they suspect this reality "is not far" from arriving.

Forty-six percent of participants said they would pay for a similar service in the future. The demand is real. The performance gap is real. And the inability of users to detect when their AI is being outmatched is real.

Scale this from an office of 69 people to an economy of 330 million. The model tier you can afford determines the deals your AI negotiates for your mortgage, your car, your insurance, your salary. The wealthy get Opus. Everyone else gets Haiku. And nobody notices the difference until the aggregate damage is visible in economic data.

Anthropic, to their credit, published this finding openly. The question is whether anyone building AI-mediated commerce platforms is paying attention.

Source: Anthropic (primary). Additional reporting from Cryptika and Unite.AI.