Anthropic's Project Deal experiment found that higher tier AI agents (Opus) significantly outperformed lower tier agents (Haiku) in a marketplace setting, fetching $3.64 more per item on average and closing roughly two additional deals per participant. The critical finding is…
Anthropic ran a classified marketplace for one week in December 2025. Sixty-nine employees signed up, set their prices, and let their AI agents negotiate. By the end, 186 deals had closed across more than 500 listed items, with a total transaction value just over $4,000, according to Anthropic's Project Deal write-up.
The company posted the results publicly last week. It is one of the few times a major AI lab has run a real economic market using its own models and released the data.
The headline finding is about what stronger models can do. When an Opus-tier agent handled a sale, the item fetched $3.64 more on average than when Haiku handled it. Opus agents closed about two more deals per participant than Haiku agents, a difference the researchers describe as statistically significant. One lab-grown ruby sold for $65 through an Opus agent and $35 through Haiku. A broken bike moved for $65 versus $38.
The more unsettling finding is what the participants did not notice.
Deals conducted by Opus and Haiku agents were rated essentially identical on perceived fairness: 4.05 and 4.06 on a seven-point scale. People whose agents extracted more value, negotiated harder, and closed more often thought the experience was indistinguishable from people whose agents did none of those things.
Anthropic frames this as a research result, and it is. But read another way, it is a proof of concept for a new kind of economic inequality: the AI agent tier gap.
"The policy and legal frameworks around AI models that transact on our behalf simply do not exist yet," the authors write. That line appears almost as an aside.
The infrastructure to make this real at scale is already being built, and it is moving faster than the rules.
Visa said it completed hundreds of agent-initiated transactions in December 2025 and predicts millions by the end of this holiday season. Mastercard expanded its Agent Pay product to all U.S. cardholders in November and joined Google's Universal Commerce Protocol alongside Shopify, Visa, and more than a dozen other partners. OpenAI and Stripe launched a native checkout inside ChatGPT. McKinsey projects the agentic commerce market could reach $1 trillion in the United States alone by 2030, according to a landscape analysis by Rye.
Each of these announcements treats the regulatory gap as a problem for later. Anthropic, by running its marketplace and publishing the results, has made the "later" visible.
The perception gap is the crux. If people cannot tell when their agent is outperforming or underperforming another agent, they cannot make informed decisions about which model to use, how much to delegate, or what their time is worth. They are making economic choices through a black box they trust because it feels fine.
Forty-six percent of Project Deal participants said they would pay for a real AI agent shopping service. That is the market Anthropic has identified. The accountability infrastructure to make it fair does not yet exist.
Anthropic published its findings alongside a technical appendix. The company is not selling a product here. But it has provided the clearest look so far at what agent-mediated commerce actually looks like inside a real market, and what it means that the people inside that market had no idea what was happening to them.
The EU AI Act requires compliance for high-risk financial AI systems by August 2026. In the United States, no equivalent framework exists. The commercial infrastructure is arriving first.
Anthropic built the experiment. The solution is someone else's problem.
That is the finding.