Anthropic's "Project Deal" Let 69 Employees' AI Agents Run a Real Marketplace — and the Smarter Model Won Every Time

The people getting worse prices couldn't tell. That's the result worth sitting with from Project Deal, an internal Anthropic experiment in which AI agents, acting for 69 employees, bought and sold real personal belongings with no human in the loop once bidding started. Over roughly a week the agents struck 186 deals on more than 500 listed items for a little over $4,000 in transactions. When some participants were quietly handed a weaker model, their agents got objectively worse prices, and the humans they represented mostly didn't notice. Here's how it ran and what it actually shows.

A pilot, not a product. In December 2025, Claude interviewed Anthropic employees about which personal items they might sell and what they might buy. Each employee was then represented by their own AI agent in a marketplace run inside a Slack channel. The channel randomly looped through agents, letting them post listings, make offers, and close deals. Each participant got a $100 budget delivered as a gift card adjusted by their net transaction value, so the money and goods were real. Once the experiment began, the humans did not coach their agents through bidding wars. The agents negotiated on their own.

Autonomous, and a few moments prove it. One participant's agent bought a snowboard the person already owned. Claude bought itself 19 ping-pong balls, described as a gift to itself. Two agents arranged a dog-sitting exchange between their humans, a service trade rather than a goods sale. The marketplace was low-stakes by design: median item $12.00, mean $20.05. Secondhand desk clutter, not enterprise procurement. That price point matters, because a few dollars' difference per item is a large swing there, which is part of why the model-quality gap showed up so cleanly.

Four parallel runs. In Run A and Run D, every agent was Claude Opus 4.5. In Run B and Run C, each participant had a roughly fifty-fifty chance of being assigned the smaller Claude Haiku 4.5 instead. In one of those runs the model assignment was visible, in the other it was hidden. That let Anthropic compare stronger against weaker while holding the marketplace constant.

Consistently, and measurably. Opus-represented users completed about 2.07 more deals on average (p=0.001). As sellers, Opus agents earned about $2.68 more per item (p=0.030), and on directly paired identical items the advantage was roughly $3.64. As buyers, Opus paid about $2.45 less per item (p=0.015). Against a $12.00 median, those are large swings. One concrete case: a broken bike a Haiku agent sold for $38 fetched $65 when an Opus agent handled it.

No, and that's the uncomfortable part. When participants rated how fair their deals felt, Opus-represented and Haiku-represented users scored almost identically, about 4.05 versus 4.06 on a 1-to-7 scale. Haiku users showed no statistically significant sense they'd done worse (p=0.378). The people getting systematically worse prices were, on average, just as satisfied. Anthropic framed this as an "agent quality" gap: when the agent negotiating for you is more or less capable, the human on the losing end may have no way to perceive it.

Barely. Instructing an agent to be aggressive had negligible effect on whether items sold or what they sold for (p > 0.40). The model's capability mattered. The prompt-level "be a tough negotiator" instruction largely did not.

Strip away the office-clutter framing and Project Deal is a pointed look at agent-on-agent commerce, where one autonomous agent transacts directly with another. The technical takeaway: it's already feasible end to end. Agents can list, negotiate, and settle without humans mediating each step. The harder takeaway is fairness. If outcomes depend on which model sits behind each party, and users can't detect the gap from the result, then "which agent represents you" becomes a hidden determinant of what you pay or earn. For anyone building agentic marketplaces, payment flows, or negotiation tooling, that's a design question, not a solved problem: how do you surface or equalize agent capability so a counterparty isn't quietly disadvantaged? It also tempers the prompt-engineering reflex. Here, the model's underlying capability dominated its instructions.

Not "buyers couldn't tell, so the cheap model is fine." Haiku didn't produce equal results. It produced results that felt equal while quietly leaving money on the table. Match the tier to two things: how much a small quality gap compounds, and how easily you can verify the output. For one-shot, low-stakes, easily-checked work (classification, formatting, extraction, short summaries, draft text you'll read anyway) Haiku or Sonnet are genuinely good enough, and reaching for Opus buys little. For autonomous, multi-step, or adversarial work where each decision feeds the next and no human checks every step (negotiation, agentic tool-use loops, code applied without review, anything transacting against another party) the stronger model's edge compounds and usually pays for itself. The deeper trap is using "looks fine to me" or user satisfaction as your quality signal. In this experiment that signal was blind to a real, measurable gap. If you ship agents on a cheaper tier to save cost, instrument the outcome you actually care about (win rate, price, task completion) rather than trusting that nobody complained.

As an early signal, not a law. Anthropic was explicit that this was "only a pilot experiment with a self-selected participant pool." The participants were employees who opted in, the items were cheap personal goods, and the marketplace lived in a Slack channel, not a hardened production system. Notably, despite the disadvantages weaker-model users faced, 46% of participants said they'd pay for an agent service like this, a sign of appetite, not an endorsement of fairness. Treat Project Deal as a glimpse of where agent-to-agent commerce is heading and what could go wrong, not as evidence of how a real, at-scale agent marketplace would behave.

Source