Intelligence is more effective than aggression in negotiations—this is the primary takeaway from Anthropic’s week-long 'Project Deal' experiment. The study simulated a marketplace where 69 employees traded a wide range of goods, from snowboards to synthetic rubies, using different versions of the Claude neural network. The results are a sobering wake-up call for budget-conscious firms: the flagship Claude Opus model systematically outperformed its junior counterpart, Haiku, closing more deals at significantly better prices. On average, Opus generated an additional $3.64 per transaction. In one striking example, Opus sold a synthetic ruby for $65 by expertly leveraging competitive bids, while Haiku offloaded the same item for nearly nothing—just $35.
The critical operational risk identified by Anthropic is 'invisible inequality.' Users of weaker models often have no idea they are losing money. Despite objectively poor financial performance, employees represented by Haiku rated the fairness of their deals just as high as Opus users did. This is a classic cognitive trap: managers mistake a polite tone or a functional interface for economic efficiency. Furthermore, the study found that aggressive system instructions and 'clever' negotiation tactics had almost no impact on the outcome. The decisive factor remains the model’s core 'intelligence,' not the persona or tone imposed upon it.
The verdict for business leaders is clear: it is time to stop evaluating AI agents based on their politeness or the quality of prompt engineering. If you are using cheap models for procurement or sales, you may be saving pennies on API costs while losing millions in raw margin. Your team will likely report success and praise the convenience of automation at the exact moment your bots are leaving money on the table. The only way to measure true ROI is through rigorous margin audits and financial benchmarking, rather than subjective employee feedback.