In a 500-day startup survival test, most AI models went bankrupt, with only three finishing above their starting capital. The test, called CEO-Bench, simulates running a fictional software company and requires strategic decision-making and long-term planning. This is significant because it highlights the limitations of current AI models in handling complex, real-world tasks with uncertain outcomes and delayed feedback. Why it matters: the results underscore the need for AI to develop more strategic and adaptive capabilities to succeed in dynamic, high-stakes environments.