Anthropic ran an AI marketplace where bots haggled over real goods — and the results reveal a hidden inequality

Anthropic quietly ran an experiment in December 2025 where AI agents — not humans — negotiated the buying and selling of real goods with real money. Sixty-nine Anthropic employees each received a $100 gift card budget and traded items with colleagues entirely through AI intermediaries. The experiment, called Project Deal, produced 186 completed transactions totaling just over $4,000, and its findings raise uncomfortable questions about who benefits when AI starts doing our shopping.

The setup

Each participant had an AI agent act on their behalf — either as buyer, seller, or both. Anthropic ran four parallel marketplaces simultaneously: one "live" market where all deals were settled after the experiment, and three separate research environments. According to the Anthropic Project Deal write-up, participants were randomly assigned either Claude Opus 4.5 or Claude Haiku 4.5 as their agent — a 50/50 split — but nobody was told which model they had.

The gap in outcomes was measurable and consistent. When the same item sold once through an Opus agent and once through a Haiku agent, Opus pulled in $3.64 more per item on average. Opus users also completed roughly two more deals overall. Neither group noticed the difference: participants rated marketplace fairness at a neutral 4 out of 7, regardless of which model was working for them.

Why it matters

The perception gap is the key finding. Users with weaker agents accepted worse outcomes without realizing it — a quiet redistribution of value that, per The Decoder analysis, participants had no way to detect in real time. Anthropic itself flags this as a concern: the company explicitly warns that policy and legal frameworks for AI agents transacting on humans' behalf don't yet exist.

That regulatory gap is already attracting attention. The FTC has been watching agentic AI closely, and with AWS reportedly planning an AI agent marketplace launch, the stakes are rising fast. Despite the unknowns, 46% of Project Deal participants said they would pay for a similar service in the future — a sign that consumer appetite for AI-handled commerce is real, even when the risks aren't fully understood.

The bigger picture

Project Deal is a proof of concept, not a product. But it sketches out a future where the quality of AI you can afford determines the deals you get — with no visible signal to tell you you're losing. Anthropic's own framing treats this as an open problem worth solving before deployment at scale, not after.