AI agent negotiation has now been measured once, in a small pilot: when two otherwise identical agents negotiate, the stronger model gets a better price, and the people behind the weaker agent do not notice. Anthropic published the experiment, called Project Deal, on April 24, 2026. This piece reads it from the seller’s side and stays inside what the page actually says.

What Project Deal was

Per Anthropic’s write-up, the company recruited 69 employees, gave each a $100 budget, and let Claude agents buy and sell on their behalf. The 69 agents “struck 186 deals across over 500 listed items” for “a total transaction value of just over $4,000”.

Anthropic ran four independent versions of the marketplace at once. In two of them (Run A and Run D) every agent was based on Claude Opus 4.5. In the other two, agents were a mix of models, which is where the comparison comes from.

The numbers that matter

  • “When an item was sold by Opus instead of Haiku, it went for $3.64 more on average.”
  • As a seller, Opus “extracts $2.68 more on average for the same item, and as a buyer pays $2.45 less.”
  • One example, quoted from the page: “Same broken folding bike. Same buyer. Same seller. Haiku sold it for $38. Opus got $65.”
  • “Perceived fairness was essentially identical for deals conducted by either model.”

The last line is the one with consequences. The weaker agent cost its user money, and nobody could tell from how the deals felt.

The caveats are Anthropic’s own

Anthropic calls it “a pilot experiment with a self-selected participant pool”, and says the experiment “wasn’t designed to dive deep into the dynamics at play here”. The items were second-hand goods traded between colleagues, not a retail catalog. Nothing in it measures a real store, a real payment rail or a shopper outside the company.

So treat it as a first data point that quality differences between agents move prices, not as a forecast.

What it implies for a merchant (our inference, not Anthropic’s)

Project Deal says nothing about product pages. The merchant angle below is our reading of it.

  • Your counterparty’s quality varies and you cannot see it. If a shopper’s agent is weak, it may pay more; if strong, less. Fixed, clearly stated prices give both a clean comparison and remove the room for either to be misled.
  • Negotiation needs facts to negotiate over. An agent can only weigh price, availability, shipping and returns if they are readable. A store that exposes them as structured fields (what agents actually parse) gives a good agent something to work with, and a weak one fewer ways to get it wrong.
  • Where a gap hides, trust rules matter. Because users did not perceive the difference, the safeguards will have to sit in the protocol and in the merchant’s published terms (return policy, price validity), not in anyone’s sense of “that felt fair”.

None of this is required by any spec today. It is a reason to care about the boring fields, not a new checklist item.

FAQ

What was Anthropic’s Project Deal?

Project Deal was a pilot in which 69 Anthropic employees each received a $100 budget and had Claude agents buy and sell items for them. The agents made 186 deals worth just over $4,000 in total, across four parallel marketplace runs. Anthropic published the results on April 24, 2026.

Did a better AI model really get better prices?

In Anthropic’s mixed-model runs, yes. Items sold by an Opus agent went for $3.64 more on average than when a Haiku agent sold them, and Opus buyers paid $2.45 less. Anthropic itself notes the pilot used a self-selected group and was not designed to explain the dynamics.

Can shoppers tell when their agent is negotiating badly?

In this pilot they could not. Anthropic reports that perceived fairness was essentially identical for deals done by either model, even though the prices differed. That is one small experiment, so it is a signal to watch, not an established rule.

Sources

Agents will negotiate over whatever a store makes readable, which is why AgentReady audits the transaction fields rather than the front page.