Princeton's CEO-Bench: Most AI Models Go Bankrupt as CEOs

Princeton University released a new test called CEO-Bench. It simulates a startup company to see how well AI models can act as CEOs. Each model had to run a startup for 500 days with $1 million in starting cash, making decisions on pricing, budgets, and competition.

Most models went bankrupt. The best result came from Claude Fable 5, which ended with $47 million in cash. But Grok 4.20 only lasted 28 days on average. Claude Opus 4.8 and GPT-5.5 also did well, but many others like DeepSeek V4 Pro and Gemini 3 Flash all failed.

The test focuses on handling uncertainty, finding information in noisy data, adapting to changes, and juggling multiple goals. The AI agents could use 34 tools and check 19 business databases, but they couldn't see customer preferences directly. They had to guess from things like subscriptions and complaints. This shows how current AI struggles with complex business decisions.