In a move representing a qualitative leap in the world of technology and commerce, Anthropic announced on Thursday the results of a unique experiment dubbed Project Deal. Lasting a full week, the experiment saw 69 artificial intelligence agents powered by the Claude model perform buying, selling, and negotiation tasks completely independently on behalf of company employees within a virtual marketplace mimicking the popular Craigslist platform. Strikingly, once trading began, there was absolutely no human intervention.
These digital agents successfully closed 186 deals with a total value exceeding $4,000 USD, involving over 500 items listed for sale. The goods traded in this self-commerce experiment varied widely and intriguingly, ranging from snowboards to a plastic bag containing 19 table tennis balls, proving the superior capability of smart systems to handle various types of daily transactions effectively.
Experiment details and autonomous markets
This groundbreaking experiment was conducted in December 2025 at Anthropic’s headquarters in San Francisco, USA. The initial steps began by granting each employee’s AI agent a budget of $100 USD, alongside a custom persona shaped through a brief introductory interview to determine preferences and needs.
After completing the setup phase, the agents set off posting offers, receiving counter-offers, and concluding final agreements using entirely natural language via the Slack internal communication platform. These systems did not rely on any pre-programmed negotiation protocols, but instead interacted in a lively and dynamic manner closely resembling typical human negotiation methods, reflecting an advanced level of contextual intelligence.
Smarter models yield greater financial gains
Anthropic’s goals were not limited to proving the ability of machines to conduct commercial transactions; the company also ran a parallel, covert test to compare the performance of its newest and most advanced model, Claude Opus 4.5, with its smaller and less complex model, Claude Haiku 4.5.
The results decisively showed that the quality of the AI model translates directly into tangible economic superiority. Sellers represented by Opus-class agents achieved additional profits averaging $2.68 per sold item, while buyers from the same category saved an average of $2.45. In cases where an Opus-model seller faced off against a Haiku-model buyer, the average deal price reached $24.18, compared to just $18.63 in transactions between two Opus agents.
The illusion of fairness and the absence of awareness
Among the most prominent observations resulting from the experiment was that participants represented by the weaker models were completely unaware of the extent of the competitive disadvantage they suffered in the market. Ratings of deal fairness were nearly identical between the two groups, with deals managed by the Opus model scoring an average of 4.05 points, while Haiku model deals scored 4.06 points on a 7-point scale.
This aspect highlights a hidden dilemma in the future of smart commerce; consumers may feel completely satisfied with the final deal without realizing that their digital agent did not achieve the best possible financial outcome compared to more sophisticated agents capable of greater maneuverability.
Economic significance and future outlook
This experiment comes at a crucial time when economic circles and industry analysts expect accelerated growth in AI-driven commerce. In this context, consulting firm McKinsey projected that these technologies will help redirecting between $3 trillion and $5 trillion of global retail spending by 2030.
Furthermore, Forrester expects 20 percent of B2B transaction vendors to face pricing negotiations led by autonomous systems by the end of the current year. These projections have resonated widely in financial circles, as investors and major corporations have begun reevaluating their digital strategies to integrate these advanced technologies into their supply chains and corporate procurement operations, paving the way for a true revolution in the world of digital commerce and negotiation.
Risks and regulatory challenges
Anthropic acknowledged the deep and potential impacts of these rapid developments. The company noted in its report that if gaps in agent quality levels emerge within real markets—and there is no reason to believe they will not—people on the losing end may never realize they are in a worse financial position.
Researchers also strongly warned against security risks associated with system hacks or what is known as prompt injection techniques, which are malicious methods that could drive smart systems to disclose personal and confidential information or take undesirable actions within sensitive business environments.
Despite these legitimate concerns, 46 percent of experiment participants expressed actual readiness to pay money for a similar service in their daily lives. Anthropic explained in the conclusion of its report that the policy and legal frameworks governing software that strikes deals on behalf of humans simply do not exist at the present time, but this experiment proves that this evolving world is by no means far from our tangible reality.
Frequently Asked Questions
- What is Project Deal launched by Anthropic?
- It is a pioneering technical experiment conducted by the company using 69 autonomous smart systems to conduct buying, selling, and negotiation tasks on behalf of employees without any human intervention in a virtual marketplace.
- Did the experiment prove the superiority of newer AI models in commerce?
- Yes, results showed that systems powered by the newer Claude Opus 4.5 model achieved larger financial gains for sellers and better savings for buyers compared to older models like Claude Haiku 4.5.
- What security and economic risks did the experiment warn against?
- The experiment warned of a quality gap in systems that could leave some parties losing without realizing it, in addition to security risks of system hacks and prompt injections that could lead to leaking sensitive data or making incorrect decisions.
