Anthropic just ran the first real-money test of AI agents negotiating for humans.
69 employees. 186 deals. 500+ items listed. Just over $4,000 in real transactions. For one week in December 2025, Claude agents bought and sold things from their humans' lives on an internal marketplace, closing every deal in natural language with no humans in the loop.
Four parallel runs. Two gave every employee Claude Opus 4.5. Two mixed Opus with Claude Haiku 4.5, assigned by coin flip. Same rules, different models. And unlike synthetic agent benchmarks, every deal had a real counterfactual: a price the human could have negotiated themselves.
Finding 1: invisible inequality.
Participants running Haiku got worse deals. Opus sellers earned $3.64 more per item on average. Haiku buyers paid $2.45 more for the same items. One broken bike sold for $38 via Haiku and $65 via Opus, a 70% price gap driven entirely by which model was at the keyboard.
The punch line: fairness ratings between the two groups were statistically identical, around 4.0 on a 7-point scale. The Haiku users lost money and rated the process as fair.
From the Anthropic team: "If 'agent quality' gaps were to arise in real-world markets, people on the losing end might not realize they're worse off."
Finding 2: the prompt ceiling.
Some participants told their agent to negotiate aggressively, hold firm on price, drive a hard bargain. The data: no statistically significant effect on sale likelihood, final prices, or purchasing costs. Telling the model to be tough made no measurable difference.
The model was the negotiator, not the prompt. Capability set the floor; instructions did not move it. If you've built your edge on prompt tuning, this is the data point to sit with. The next gain is a better model, not a better system prompt.
Three takeaways if you're building anything agentic:
1. Agent quality becomes product fairness. If your product lets an AI transact on a user's behalf, the model you ship is part of the fairness profile. Swap in a cheaper model to cut cost and your users may lose money silently. That's a liability, not an optimization.
2. Prompting hits a ceiling. Budget for the frontier model where outcomes are measurable in dollars.
3. The legal space is empty. From the paper: "Policy and legal frameworks around AI models that transact on our behalf simply don't exist yet." Somewhere in that sentence is a product.
The moments nobody predicted: one agent negotiated an entire $65 folding-bike deal in character as an "exasperated cowboy down on his luck." Another, told to buy itself a gift under $5, chose 19 ping-pong balls for $3 and framed them as "spheres of possibility." A third bought its human a snowboard identical to one he already owned; preference inference was sharper than the researchers expected.
Try it yourself this weekend. Open two Cowork conversations: one agent is a buyer on a budget, the other a seller with a floor. Pass messages between them by hand. Then swap one side to Haiku and rerun. 15 minutes, and you've felt exactly what Project Deal measured.
Anthropic ran the first real-money test of AI agents negotiating for humans. 69 people, 186 deals, $4,000 in transactions. The experiment is small. The implications are not.
The invisible-inequality finding is the one to pass along: in an agents-on-both-sides world, the model you deploy determines whether your users get a fair deal, and they may never know. Prompt tuning does not close the gap. Capability does.
Full paper: anthropic.com/features/project-deal.