Anthropic conducted an internal experiment Project Deal - a test platform where AI agents on behalf of employees bought and sold goods without the participation of humans. Each of the 69 participants was allocated a $100 budget, and Claude pre-interviewed them to form personal extensions. The market was expanded to Slack, with agents posting ads, negotiating, and closing deals. Total 186 transactions were made on more than 500 lots worth over $4,000.

The experiment showed that the model’s quality had a noticeable effect on the result. Users with an advanced Claude Opus 4.5 made on average two more deals than owners of a weaker Haiku 4.5, and sold the same goods at a higher price - the difference averaged $3.64, and in some cases was short. At the same time, attempts to give an agent aggressive or friendly instructions did not yield a statistically significant commercial benefit.

Separately, Anthropic noted risks: a less powerful model could imperceptibly worsen the owner’s position in the trading.