When Agents Negotiate: What Anthropic's Project Deal Tells Us About Open Commerce Protocols

Anthropic published an experiment that should reshape every agentic commerce conversation. Opus agents extract $2.68 more per sale while user fairness ratings stay identical. Here's why open commerce protocols are the only credible response.

On 29 April 2026, Anthropic published an internal experiment they call Project Deal. Sixty-nine employees, each paired with a Claude agent, traded on a live internal marketplace. The numbers buried in the report are the kind that should redirect a product roadmap. Opus agents extracted $2.68 more per sale than agents running on smaller models. Opus buyers paid $2.45 less for equivalent goods. Fairness ratings were statistically identical — 4.05 vs 4.06 out of 7. Users could not feel the gap. Anthropic's own framing is worth quoting verbatim: "Agent quality gaps could quietly entrench inequality in agentic markets." That is a research lab telling the industry, on the record, that the marketplaces being built right now will have systematic, invisible economic stratification baked in by the model tier of each participant's agent. The hidden cost of trusting the marketplace owner Project Deal is a closed system. Anthropic could measure the gap because they ran both sides of the experiment. In a real agent commerce market — OpenAI's checkout, Google's UCP, Amazon's Rufus — the data live behind walled gardens. The marketplace owner sees the gap. The participants do not. This matters because the marketplace owner has no economic incentive to disclose tier gaps. A platform that takes a cut of every transaction profits from price dispersion, not from price flattening. The buyer who paid $2.45 too much will never know they paid $2.45 too much. The seller who left $2.68 on the table will never see the comparison. The only way to surface this asymmetry is structural transparency — and structural transparency is incompatible with proprietary infrastructure. What "structural" means in practice There are three levels at which an agent commerce protocol can disclose quality gaps. Nexbid has a draft specification covering all three: Mandate-embedded disclosure (Enterprise tier). Every signed agent mandate carries a verifiable claim about which model tier executed the negotiation. A merchant integrating with Nexbid sees model_tier: "opus-4-7" or model_tier: "haiku-4-5" as part of the cryptographic envelope. Procurement teams can require parity in their contracts. User-side surfacing. Consumer-facing wallets render the tier as part of the purchase confirmation. Not as fine print — as a first-class field. "Your agent: Claude Opus. Counterparty's agent: Claude Haiku." The user sees the asymmetry the same way they see the price. Public audit trail. Aggregated, anonymised tier-pairing statistics published quarterly. Not per-user data — distribution data. In the past 90 days, …

Author

Holger von Ellerts

Published

2026-04-29

Topics

Research, Protocol Commerce, Agentic, Open Source