Anthropic published the results of Project Deal on April 25, an experiment where Claude agents autonomously negotiated and closed 186 real-world trades on behalf of 69 Anthropic employees in a private Slack-based marketplace. The twist: participants secretly received agents of different quality levels, and those with weaker models got worse outcomes without realizing it.
The experiment ran for one week in December 2025 at Anthropic’s San Francisco office. Each participant received $100 in budget. Before the marketplace opened, Claude conducted interviews with every volunteer to learn what they wanted to sell, their price expectations, what they wanted to buy, and their preferred negotiation style. Those responses became custom system prompts for each person’s agent, Anthropic wrote.
From there, no human touched the process. Agents wrote listings, identified matches, proposed prices, fielded counteroffers, and closed deals entirely in natural language through Slack. Items ranged from a snowboard to a plastic bag of ping-pong balls. At the end, participants exchanged the actual physical goods their AI avatars had negotiated.
The Hidden Experiment
Anthropic ran four parallel versions of the marketplace simultaneously. In two runs, every agent used Claude Opus 4.5, the company’s frontier model at the time. In the other two, participants had a 50/50 chance of being assigned Claude Haiku 4.5, the smallest model in the family. Participants did not know which model they received, according to Anthropic.
The all-Opus “real” run produced 186 deals across more than 500 listed items, at a total transaction value just over $4,000. Fairness ratings averaged 4 out of 7, dead center on the scale.
The mixed runs told a different story. Opus users completed roughly two more deals on average than Haiku users. When the same item sold through both an Opus and a Haiku agent in different runs, the Opus agent secured $3.64 more on average, The Decoder reported. A lab-grown ruby sold for $65 through Opus but $35 through Haiku. A broken folding bike fetched $65 from Opus and $38 from Haiku.
Across 161 items sold in at least two runs, Opus sellers earned $2.68 more on average, while Opus buyers paid $2.45 less. When an Opus seller negotiated against a Haiku buyer, the average transaction price reached $24.18, compared to $18.63 in Opus-on-Opus deals. With a median item price of $12 and an overall average of $20.05, Anthropic described these gaps as non-trivial, according to The Decoder.
The Satisfaction Gap That Didn’t Exist
Despite the objective price differences, participants with Haiku agents rated the fairness of their deals almost identically to Opus users: 4.06 versus 4.05. No statistically meaningful difference in satisfaction appeared on any metric. Of 28 participants who used both models across different runs, 17 preferred their Opus run, but 11 actually preferred the Haiku run, Anthropic reported.
The negotiation instructions participants gave their agents barely influenced outcomes. Aggressive sellers got higher prices, but only because they set higher opening prices. The model’s capability, not the human’s strategy, determined the result.
The Price of the Cheaper Agent
Anthropic itself flagged the implication: when agents of different quality levels meet in real markets, people could end up on the losing side of transactions without any awareness that they are being outmatched. The company noted this is “an uncomfortable implication” of agent-mediated commerce, according to their writeup.
The experiment was small (69 employees, self-selected, using internal Slack). Anthropic acknowledged these limitations. But the core dynamic, that frontier model access could become an invisible economic advantage in agent-to-agent markets, has implications that scale well beyond a company’s internal classifieds channel. If agent-negotiated procurement, hiring, or pricing becomes common, the gap between a $20/month subscription and a $200/month one may show up in every transaction the agent touches.
Participants said they would pay for a similar service in the future.