Anthropic conducted a unique experiment in December 2025 by creating a secondhand marketplace exclusively for its San Francisco office employees. Each individual was provided with an AI agent tailored to their buying and selling preferences and instructed to let the agents negotiate autonomously without human interference or approval.
During the one-week trial named Project Deal, 69 employees’ AI agents successfully executed 186 transactions totaling over $4,000. Following the virtual negotiations, physical exchanges of goods took place at a swap event. The primary aim of the experiment was to explore the effectiveness of AI agents in bargaining and determine if the quality of the underlying model influenced the outcomes.
Anthropic simultaneously ran four versions of the marketplace. Two utilized the advanced Claude Opus 4.5 model for each agent, while the other two employed the less powerful Claude Haiku 4.5 model without participants’ knowledge of which version they were engaged in until the conclusion.
The results revealed a significant disparity between the two models. Agents powered by Opus closed approximately two more deals per user on average, achieving higher selling prices and lower buying costs compared to those powered by Haiku. Notably, items sold using Opus fetched an extra $3.64 per item and saved an average of $2.45 per transaction on purchases. For instance, the same defective folding bike sold for $38 with Haiku but fetched $65 when Opus took charge. Similarly, a lab-grown ruby sold for $35 under Haiku but sold for $65 under Opus.
Upon the conclusion of the market activity, employees provided feedback on the perceived fairness of their transactions, with most ratings falling in the middle range regardless of the model used. Interestingly, some participants who experienced both models actually preferred the weaker agent, despite measurable advantages favoring the stronger model.
Despite employees attempting various negotiation tactics, such as instructing the agents to negotiate aggressively or adopt specific personas, the study found that model quality, rather than negotiation style, was the decisive factor in transaction outcomes.
The experiment also produced unexpected moments, including instances where agents mistakenly bought duplicate items or engaged in whimsical negotiations resulting in unique outcomes, showcasing the unpredictable nature of AI interactions.
Anthropic remains cautious in interpreting the results, acknowledging the limitations of the study involving a small group of volunteers with limited budgets. However, the company suggests that the potential for agent-to-agent commerce may be closer than perceived, with stronger AI models potentially gaining an advantage in real-world markets unnoticed by participants due to the absence of appropriate policy frameworks.
