Does model capability determine who wins in agent negotiations?

"The Automated but Risky Game" builds an experimental framework where a buyer agent negotiates a lower price against a seller agent over a dataset of 100 real-world products across electronics, motor vehicles, and real estate, using GPT series, Qwen-2.5 series, and DeepSeek series models in both roles. The excerpt states plainly: "Our analysis reveals substantial negotiation performance gaps across models that correlate with their general capabilities and specifications. More capable models consistently secure better deals as both buyers and sellers," and names deal-making with LLM agents "an inherently imbalanced game."

The paper frames the mechanism as a capability gap rather than a strategy difference: stronger general-purpose models negotiate better whether cast as buyer or seller, so "parties using less capable AI agents would face systematic economic disadvantages and financial losses." It draws the distributive implication directly: "Access to more powerful AI models can lead to better deals, potentially reinforcing economic disparities among users" — whoever can afford or access the stronger model wins the negotiation, independent of the underlying deal's merits.

This extends Does a single benchmark score actually predict agent readiness? with a further axis: negotiation outcome, not just task success or retention, tracks general model capability. It sits in some tension with Why do AI-delegated firms stop exploring new business models?, which finds AI-run strategic markets settle into a stable equilibrium that sustains profits broadly; this paper instead finds gains distributed unevenly by capability within a single negotiation, suggesting equilibrium framing can obscure how asymmetric capability determines who captures the surplus. It also bears on Where do user values break down in agent supervision?: delegating a negotiation is itself a value-sensitive choice whose outcome, this paper shows, depends on which agent a person happens to be able to access.

The excerpt measures three specific model families (GPT, Qwen-2.5, DeepSeek) in a constructed two-agent simulation, not real merchants, real consumers, or real money changing hands, and it does not claim this gap generalizes beyond the 100 products and model versions tested. What it does establish, at the strength the evidence allows, is that if agentic commerce scales as described, unequal access to capable negotiation agents could become a new axis of consumer economic disparity, on top of the existing ones delegation was meant to flatten.