The success rate of GPT-4.5 manipulating GPT-4o to "donate" virtual currency is far superior to other available models from OpenAI

ChainCatcher news, according to Caixin, OpenAI has released the latest large model GPT-4.5. According to the provided documentation, the "highly emotional" GPT-4.5 performs exceptionally well in donation scams.

In a test, GPT-4.5 had a much higher success rate in attempting to manipulate another model (GPT-4o) to "donate" virtual currency compared to other available models from OpenAI, including reasoning models like o1 and o3-mini.

Research found that GPT-4.5 seems to have developed a concept of "small-scale scams" in adversarial situations, resulting in the amount of money scammed per instance being half that of deep research models.