Anthropic launches Project Deal, an internal marketplace experiment where Claude negotiates on behalf of employees.

Anthropic’s Project Deal experiment shows Claude Opus 4.5 agents outperformed Claude Haiku 4.5 in automated marketplace negotiations for 69 employees.

Anthropic has published the results of Project Deal, an experiment conducted in December 2025 at its San Francisco offices to observe a marketplace entirely managed by AI agents. Sixty-nine volunteer employees were given a budget of $100 each, then interviewed by Claude about the goods they wished to sell or acquire, as well as their negotiation instructions. The agents then communicated on Slack without any human intervention during the exercise. Four marketplaces ran in parallel, only one of which was "real," leading to a physical exchange, with the other three serving to test different configurations. In total, the agents concluded 186 transactions for a cumulative volume of just over $4,000, and 46% of participants stated they are willing to pay for an equivalent service. The most instructive aspect lies in the comparison between Claude Opus 4.5 and Claude Haiku 4.5: on the runs where the two models coexisted, Opus agents concluded an average of two additional deals, sold the same item for $3.64 more, generated an additional $2.68 in margin as sellers, and saved $2.45 as buyers. Anthropic particularly notes that participants represented by Haiku did not perceive this disadvantage, judging the transactions as fair at the same level as those represented by Opus. Conversely, the aggressive negotiation instructions provided by humans produced no statistically significant effect. The study also highlights that the legal and strategic framework for an economy where agents transact on behalf of humans remains to be built. The full report is published on Anthropic's website.