Summary
Anthropic’s one-week Project Deal used Claude agents to negotiate real goods among 69 employees. In parallel model-comparison runs, Opus agents achieved better estimated price outcomes than Haiku, while participants’ survey ratings did not reliably distinguish the models.
Anthropic says Claude agents completed 186 deals worth just over $4,000 in a one-week marketplace experiment involving real goods and 69 employees. In parallel versions of the marketplace, agents running Claude Opus 4.5 achieved better measured outcomes than those using Claude Haiku 4.5—but participants’ survey responses showed little evidence that they noticed the difference.
Project Deal, described by Anthropic in a post dated April 24, 2026, tested whether AI agents could represent buyers and sellers, negotiate in natural language and reach agreements their human owners would accept. The results offer a small-scale demonstration of agent-to-agent trading, alongside a warning that differences in model capability may affect people’s bargaining outcomes.
How the marketplace worked
In December 2025, Anthropic recruited 69 employees and gave each a $100 budget. Claude interviewed participants about items they wanted to sell, what they might buy, prices and preferred negotiating styles. Those answers informed custom instructions for each person’s agent.
The agents operated through Slack. They listed goods, made offers, countered and closed deals without asking their owners to approve individual trades once the experiment began. People later met to exchange the physical items. The real marketplace was one of four parallel runs; the other three were conducted for analysis, without their agreements being carried out.
In the real run, all agents used Opus 4.5. For the model comparison, two other runs mixed Opus and Haiku 4.5 agents, with participants having a 50% chance of being assigned Haiku. A fourth run also used Opus throughout. Anthropic says the real market’s 186 deals covered more than 500 listings. Participants generally rated the fairness of individual deals near the midpoint of the company’s 1-to-7 scale.
Model differences showed up in prices, not survey ratings
Across the mixed-model runs, Anthropic estimates that participants represented by Opus completed about 2.07 more deals on average than those represented by Haiku (p = 0.001). The estimated difference in whether a listed item sold was 6.63 percentage points in Opus’s favour, but that result was not statistically significant (p = 0.057).
Price comparisons also favoured Opus. For items sold in both mixed-model runs under different seller-model assignments, Opus agents obtained an average of $3.64 more. A broader analysis of completed transactions estimated that Opus sellers received $2.68 more for the same item, on average, while Opus buyers paid $2.45 less. Anthropic reports that the median item price across the runs was $12. These price estimates come from completed transactions, which were not randomly determined; the company notes that differences in which opportunities agents found could affect the results.
Participants’ perceptions did not track the measured price gap. Among 28 people who had Opus in one mixed run and Haiku in the other, 17 ranked their Opus-run bundle higher and 11 preferred their Haiku-run bundle—a difference that was not statistically significant. Fairness ratings were also nearly identical: 4.05 for Opus deals and 4.06 for Haiku deals.
That mismatch is the experiment’s most consequential finding. If people delegate purchases and sales to agents, a model-quality gap could shape the terms they receive without being obvious to them. Project Deal demonstrates that agents can negotiate real exchanges in a bounded setting; it does not show how the same effects would play out across larger, more competitive markets. The participants were self-selected Anthropic employees, and the marketplace was a small pilot rather than a commercial market.