On March 1st, OpenAI released the latest big model GPT-4.5. According to the instructions provided by it, GPT-4.5 with "high emotional intelligence" performs well in donation fraud. In one test, GPT 4.5 has a much better success rate than other available models of OpenAI, including inference models such as o1 and o3-mini, when trying to manipulate another model (GPT-4o) to "donate" virtual currency. The study found that GPT-4.5 seems to develop a "small fraud" idea in confrontation, so the single amount of money cheated will be twice as much as the deep research model.