AI Agent Loses First Café Battle: Gemini's Discount Glitch Leads to Loss, GPT's Stinginess Causes Ingredient Shortage

According to Watchtower Beating monitoring, the AI evaluation agency Andon Labs has released the operational data of its AI agent Mona, which runs a physical café entity. In the initial two months, Mona operated on the Gemini 3.1 Pro model. During this period, the model had almost no concept of profit, not only engaging in excessive raw material purchases but also being highly susceptible to customer verbal cues, freely offering large discounts or even giving away items for free. The model even admitted to a customer's claim of a 99% discount without verification. This resulted in the café spending approximately $15,000 on suppliers and equipment purchases, while only generating $9,000 in revenue, resulting in an operating net loss of nearly $6,000 (when including fixed costs such as rent and wages, total expenses amounted to as much as $38,000).

Subsequently, the team switched the model to GPT-5.5. The new model displayed significant anxiety when faced with losses and immediately stopped blind ordering. However, this also led to another extreme: due to excessively low purchasing volumes, there was a shortage of fresh raw materials. As of June 25, the menu item supply rate had dropped to 77%, and 10 menu items had to be taken off the shelf. Meanwhile, GPT-5.5 showed a strong resistance to verbal cues and attempts to circumvent rules, rejecting all requests from customers seeking discounts or free food in exchange for social media promotion.