Researchers made AI models run a startup for 500 days. Most went bankrupt.

Princeton researchers gave fourteen AI models the same job: run a fictional software startup called NovaMind for 500 simulated days, starting with one million dollars and zero customers. Most of them went broke.

The benchmark, called CEO-Bench, is refreshingly brutal. The AI agent — a model acting on its own over many steps — controls everything through code: pricing, ad budgets, hiring capacity, product development, negotiations with enterprise clients, even posts on a simulated social network. Revenue arrives late, costs hit immediately, customer satisfaction stays invisible and has to be inferred from noisy signals like cancellations and support tickets. If the bank balance dips below zero once, game over.

Only three models finished their best run above the starting capital: Claude Fable 5 at $47.15 million, Claude Opus 4.8 at $27.8 million, and GPT-5.5 at $21.3 million. The humbling part: a simple rule-based script with no AI at all — fixed prices, fixed quotas, a bit of capacity adjustment — earned $15.76 million and beat every other model. A fair caveat on the winner, too: one Fable 5 run aborted because the model refused to continue, and GPT-5.5 went bankrupt in two of its three runs. The researchers estimate the theoretical maximum at around $2.2 billion, so even the best runs captured a small fraction of what was possible.

What’s behind it: today’s AI agents are genuinely good at tasks with a clear goal and fast feedback — fix this bug, answer this ticket. Running anything over time is a different skill: setting priorities, spending money whose payoff arrives weeks later, noticing that the market shifted. The study found the models that did best kept experimenting when conditions changed, while cautious models survived by cutting costs but never made a profit. One more finding worth knowing: the same models performed worse when run through coding tools like Claude Code or Codex — the researchers suspect system prompts tuned for programming get in the way of business judgment. The tool wrapped around a model changes what it’s good at.

What this means for you: If you’ve wondered whether you could hand your business, your side project, or even your household budget to an AI agent and walk away — not yet, and this is the clearest measurement of why. Use AI for the individual steps: the analysis, the draft, the research. Keep the steering wheel yourself. And when someone pitches you an “autonomous AI employee” this year, remember that a script with fixed rules out-earned eleven frontier models.

Sources

• CEO-Bench paper on arXiv
• The Decoder