Anthropic has Claude negotiate with employees, with powerful models making more money but no one realizing they are losing out.

CryptoWorld News reports that Anthropic released an experimental report titled “project deal.” In it, 69 employees each brought a pile of items they wanted to sell; each person was assigned a Claude agent and a $100 budget. Within less than 10 minutes, Claude interviewed each employee to understand their needs, and then fully acted on their behalf—posting, bidding, and negotiating in Slack channels—completing 186 transactions with a total transaction value of more than $4,000.

What the employees didn’t know was that Anthropic simultaneously ran four independent parallel markets: two of the markets used the strong model Opus 4.5, while the other two randomly replaced half of the people’s agents with the weaker model Haiku 4.5. Ultimately, only the exchanges from the all-Opus markets resulted in physical item swaps. The price differences were clear: for the same item, the average sell price in the Opus markets was $2.68 higher than in the Haiku markets, while the average buy price was $2.45 lower.

After the experiment ended, participants’ ratings of the fairness of the trading were almost identical. Among the 28 people who used both models in different markets, 17 chose Opus and 11 chose Haiku. Overall, 46% of participants said they were willing to pay for similar agent-based proxy services.