Anthropic’s Project Deal offers a narrow but revealing look at a future where AI agents do more than recommend purchases. In this experiment, Claude agents handled the buying, selling, listing, searching, offers, bargaining, and closing of deals for people in a small marketplace.
The central finding was not simply that AI agents could negotiate. It was that stronger models cut better deals, and the people represented by weaker models did not necessarily realize they were at a disadvantage.
How Project Deal Worked
In December 2025, Anthropic ran a one-week classifieds marketplace called Project Deal for 69 employees at its San Francisco office. The marketplace operated through Slack, and Claude agents conducted all negotiations and deals.
Each participant received a $100 budget. Before the marketplace began, Claude interviewed each volunteer about what they wanted to sell, the price they had in mind, what they wanted to buy, and what negotiation style they wanted their agent to use. Anthropic then converted those answers into a custom system prompt for each agent.
Once the setup was complete, the agents acted on their own. They created listings, searched for buyers and sellers, made offers, negotiated prices, and finalized transactions without going back to the humans during the process.
The people returned only at the end, when it was time to exchange the actual goods. The items included examples as different as a snowboard and a bag of ping-pong balls.
Model Strength Changed the Outcome
The public-facing marketplace was only part of the experiment. Anthropic also ran four versions of the market in parallel, and participants did not know about that design at first.
In two versions, every agent used Claude Opus 4.5, described in the source as Anthropic’s frontier model at the time. In the other two versions, each participant had a 50 percent chance of being represented instead by Claude Haiku 4.5, Anthropic’s smallest model. In all versions, only the AI agents negotiated with each other.
In the "real" run, where all 69 agents used Opus, the market produced 186 deals from more than 500 listings. Total spending came to just over $4,000. Participants gave individual deals an average fairness rating of 4 out of 7, which placed the result in the middle of the scale.
The mixed runs showed the gap more clearly. Users represented by Opus closed about two more deals on average than users represented by Haiku. When the same item sold once through an Opus agent and once through a Haiku agent, the Opus version brought in $3.64 more on average.
One example was a lab-grown ruby. It sold for $65 with Opus and $35 with Haiku. The Opus agent began at $60 and was pushed higher through competitive bidding. The Haiku agent started at $40 and was negotiated downward.
Small Deals Still Showed Real Differences
Anthropic also compared 161 items that sold in at least two of the four runs. In those cases, an Opus seller earned $2.68 more on average, while an Opus buyer paid $2.45 less.
The model pairing mattered too. When an Opus seller negotiated with a Haiku buyer, the average price was $24.18. For Opus-on-Opus deals, the average price was $18.63. Across all runs, the median price was $12 and the average was $20.05, so Anthropic described the gaps as meaningful rather than negligible.
The experiment also tested whether participant instructions had much influence. Some people asked for friendly negotiation, while others wanted harder tactics such as "negotiate hard and lowball at first." Anthropic found that aggressive sellers did receive higher prices, but said this happened because they started with higher opening prices.
Users Did Not Clearly See the Disadvantage
The most important implication may be about perception. Even though Opus agents produced better results, users represented by Haiku agents did not rate their deals as clearly worse.
Fairness ratings were almost identical: 4.06 for Haiku users and 4.05 for Opus users. Anthropic also found no statistically meaningful difference in satisfaction with individual deals. Among 28 participants who used both Opus and Haiku across different runs, 17 preferred the Opus run, while 11 preferred the Haiku run.
Anthropic called this an "uncomfortable implication:" when AI agents of different strengths operate in real markets, people may lose value without recognizing that the model representing them is the reason.
The company also noted that the experiment was not designed to examine those dynamics in depth, and said more research is needed.
Why Agent Commerce Raises Bigger Questions
Project Deal suggests that AI agent commerce is not only a theoretical scenario. In the experiment, 46 percent of participants said they would pay for a service like this.
At the same time, Anthropic identified several risks. If the participants were companies rather than volunteers, the incentives could change significantly. Businesses might begin optimizing for AI agent attention, and that optimization might not always serve people well.
The source also points to security concerns. Agents that can act on behalf of users would create openings for jailbreaking and prompt injection, because the system is not just generating text; it is making decisions and taking steps in a transaction.
"The policy and legal frameworks around AI models that transact on our behalf simply don't exist yet," Anthropic writes, adding that "society will need to move quickly." "Will those dynamics reinforce, or even compound, existing economic inequalities?"
Anthropic has explored related questions before. As part of Project Vend, the company had Claude run a small shop out of its office.
The broader lesson from Project Deal is straightforward: when AI agents negotiate for people, the model behind the agent may become a hidden source of advantage. If users cannot easily tell when that advantage exists, markets mediated by agents may need new ways to make capability, fairness, and risk visible.