Anthropic ran an internal experiment in December 2025 in which AI agents based on its Claude models independently bought and sold physical goods on behalf of employees, with no human involvement during negotiations. The San Francisco-based AI company detailed the project, called “Project Deal,” in a blog post published in April 2026.
The week-long pilot recruited 69 employees at Anthropic’s San Francisco office and gave each participant a $100 budget to use in a custom Slack-based marketplace. Claude first interviewed each participant to learn what items they wanted to sell, what they hoped to buy, and how they preferred to negotiate. Each person was then assigned a personalised Claude agent with specific pricing and behavioural instructions.
Once the marketplace launched, agents operated without human approval — posting listings, identifying matches, making offers, negotiating prices, and finalising deals independently. When agreements were reached, agents generated deal terms that employees later executed in person by physically exchanging the agreed goods. The experiment produced 186 completed deals across more than 500 listed items, with a total transaction value of just over $4,000. Anthropic noted that participants expressed willingness to pay for a similar service in the future.
Anthropic also used the experiment to test how AI model quality affects outcomes. It ran four parallel versions of the marketplace — one real and three observational — using either the flagship Opus 4.5 model or a mix of Opus and the smaller Haiku 4.5 model. Participants were not told which version they were in until after the experiment ended.
The results showed measurable differences. Users represented by Opus completed roughly two more deals on average than those using Haiku. When the same item was sold by each model in separate runs, the Opus agent obtained $3.64 more on average — a broken folding bike, for example, fetched $38 via Haiku but $65 via Opus. Overall, Opus earned $2.68 more as a seller and paid $2.45 less as a buyer.
Anthropic said the findings raise concerns about “agent quality” disparities in real-world AI marketplaces, where users represented by less capable models could be at a financial disadvantage without knowing it.
Source: mint – technology