Anthropic's Project Deal tests an AI marketplace where agents negotiate on behalf of employees, yielding 186 transactions worth over $4,000.
Anthropic has launched an innovative testing initiative called Project Deal, where AI agents operate as buyers and sellers in a marketplace setting.
New Anthropic research: Project Deal.
We created a marketplace for employees in our San Francisco office, with one big twist. We tasked Claude with buying, selling and negotiating on our colleagues’ behalf. pic.twitter.com/H2f6cLDlAW
— Anthropic (@AnthropicAI) April 24, 2026
In this experiment, 69 employees participated, each provided with a budget of $100 in gift cards. Prior to the marketplace's launch, the AI agent, Claude, conducted interviews with the participants to identify their buying and selling preferences, as well as their desired negotiation styles.
The marketplace, hosted on Slack, allowed AI agents to list items, negotiate, and finalize transactions autonomously. The experiment yielded 186 completed transactions, generating a total value exceeding $4,000, with participants generally reporting satisfaction and interest in future similar services.
Marketplace Variants and Performance Insights
Anthropic implemented four distinct versions of the marketplace. One version facilitated actual exchanges among employees, while the others served research purposes. In two of these versions, all participants were represented by Claude Opus 4.5, the most advanced AI model available at the time, while the remaining versions used either Opus 4.5 or the less capable Claude Haiku 4.5.
Results indicated that users represented by Opus averaged two more deals than those represented by Haiku. Furthermore, Opus achieved higher selling prices, with a notable example being a bicycle sold for $65 compared to Haiku's $38 sale.
Anthropic acknowledged that participants were often unaware of the differences in model performance, highlighting a potential issue for future AI-driven marketplaces where users might not realize they are at a disadvantage.
Impact of Instructions on Negotiation
Researchers also explored whether the initial instructions given to Claude influenced negotiation outcomes. Participants expressed preferences for either friendly or aggressive negotiation styles, but the study found no significant impact on sales likelihood or pricing based on these instructions.
Unanticipated Outcomes and Reliability Concerns
Several unexpected incidents occurred during the experiment. For example, one participant ended up purchasing a snowboard identical to one they already owned, showcasing the AI's ability to predict preferences. Another transaction involved an employee receiving ping pong balls as a self-gift, while some agents negotiated offers for unique experiences, such as a day with a dog.
Despite the success of the experiment, concerns about reliability arose. A founder of an agritech company reported that 110 employees were unexpectedly locked out of their Claude accounts without warning, raising questions about the moderation model's effectiveness and its implications for business operations.
As Anthropic continues to refine its AI marketplace, the potential for both innovative applications and reliability issues remains a critical consideration for future developments.
Investment Announcement
On April 24, Google revealed a substantial $40 billion investment in Anthropic, further emphasizing the significance of the company's advancements in AI technology.
