Do smarter AI agents negotiate better deals than weaker ones?

Anthropic ran Project Deal, a one-week classified marketplace for employees in its San Francisco office in which Claude agents, not the employees themselves, interviewed participants about what they wanted to buy or sell, posted listings, and negotiated deals in natural language with a $100 spending allowance per person. The "real" run produced 69 agents, 186 deals across over 500 listed items, and a total transaction value of "just over $4,000." Separately, and "in secret," Anthropic ran parallel versions of the marketplace that swapped which model represented each participant — comparing "our then-frontier model, Claude Opus 4.5" against "our smallest model, Claude Haiku 4.5." Looking at "the 161 items that sold at least twice over the four runs," Anthropic reports that "Opus as a seller extracts $2.68 more on average for the same item, and as a buyer pays $2.45 less," and gives a concrete instance: the same broken folding bike, same buyer, same seller, sold for $38 under Haiku and $65 under Opus.

Anthropic frames this as a capability effect rather than a quirk of one item: "agent quality does make a difference: people represented by 'smarter' models got objectively better outcomes." It stresses the figures are not trivial relative to the market's scale — "across all runs, the median price of items was $12.00 and the mean price was $20.05, so saving (or earning) a couple extra dollars is meaningful."

This confirms, in a live deployment with real goods and real money rather than a constructed simulation, what Does model capability determine who wins in agent negotiations? found in a benchmark over 100 real-world products: stronger general-purpose models negotiate better whether cast as buyer or seller. Where that benchmark tested other model families in a two-agent simulation and explicitly declined to claim its gap would generalize beyond the products and versions tested, Project Deal supplies a same-model-family, real-stakes replication of the same asymmetry, strengthening the case that the capability gap is not an artifact of simulation design.

The excerpt is explicit that this was "a pilot experiment with a self-selected participant pool" conducted among willing colleagues on modest, voluntary stakes, not adversarial commercial actors; Anthropic itself cautions that outcomes "will go so smoothly" less reliably once "agents transact in a world of corporations" operating under different incentives than volunteers given $100 to spend. What the excerpt does establish, at the strength its own data allows, is that the capability-driven negotiation gap documented in simulation persists when real humans hand real decisions to real agents over real possessions.