The pursuit of bigger artificial intelligence models is giving way to a focus on autonomous agents—systems that can act, decide, and even negotiate for users in everyday scenarios. A recent study explored what happens when both sides of a negotiation—buyer and seller—use competing artificial intelligence agents. The findings point to a notable advantage for parties using stronger, more advanced agents, with those leveraging older or weaker systems routinely losing out in financial deals.
Researchers ran experiments with major language models acting as buyers and sellers in simulated negotiations involving electronics, vehicles, and real estate. The results showed a clear hierarchy: OpenAI’s newer ChatGPT-o3 and GPT-4.1 models consistently struck better deals compared to older or less capable models like GPT-3.5. DeepSeek’s latest models also performed well as sellers, while Qwen2.5 was more successful as a buyer. Not only did advanced models maximize profits or minimize spending, but some regularly outperformed others by a significant margin, mirroring real-world disparities that could emerge as artificial intelligence agents enter financial decision-making arenas.
However, the study highlighted more than just financial gaps. Even the leading models faltered at times, failing to close deals or getting entangled in endless negotiations. The researchers caution that differences in model architecture, training data, and especially model size contribute to these imbalances. They point to a bigger risk: if artificial intelligence agent negotiations become widespread, disparities in model access or sophistication could cement new forms of digital inequality, especially in domains like business procurement or finance where negotiation is common.
Experts urge caution and further research, emphasizing the simulated nature of these studies and the complexity of real markets. Some suggest that artificial intelligence agents should be evaluated not just by their best-case outcomes but also by how reliably—and safely—they handle stress and adversity, since even rare failures might result in significant systemic risk. While consumer-facing tools like Amazon´s ´Buy for Me´ avoid direct price negotiation for now, business platforms are beginning to experiment with artificial intelligence-powered sourcing assistants. For now, researchers advise using artificial intelligence as an informational resource, not a stand-in for human judgment in critical negotiations.
