In CEO-Bench, a Princeton University test that tasks AI agents with running a fictional software company for 500 simulated days, only three of fourteen models finished above their starting capital of one million dollars, and a simple rule-based system with no AI at all beat nearly all of them. The benchmark exposes how poorly today's models handle long-horizon strategy, even as they excel at short, well-defined tasks.
What CEO-Bench measures
Most AI benchmarks reward narrow skills: fixing a bug, following a policy in a chat, or completing a web workflow. These share a tidy structure, a clear goal, brief action and quick feedback. CEO-Bench targets the opposite: long chains of decisions under uncertainty, where an agent must set priorities, allocate limited resources, read noisy signals and adapt as conditions shift. The researchers call this "steering intelligence," and cite Apple's 1997 near-bankruptcy turnaround as the kind of strategic decision-making they want to test.
In the simulation, an agent runs a subscription software firm called NovaMind, starting with zero customers and one million dollars. It controls the company through a Python API with 34 tools and a 19-table database, writing its own code and SQL queries. If the balance ever drops below zero, the company is bankrupt and the run ends.
Why the test is so hard
Time and hidden information make it brutal. Revenue only arrives at billing dates, R&D takes days or weeks, and mistakes often surface late through churn or reputation damage, while costs hit immediately. The agent cannot directly see customer satisfaction or willingness to pay, and must infer them from cancellations, support tickets and a simulated social network. The world keeps moving too: competitors raise quality expectations, preferences drift, and a business cycle shifts demand.
Most models go bankrupt
Nearly every model could issue valid commands and queries, but none could sustain a coherent strategy, and many went broke before day 500. Only three cleared their starting capital on their best run:
- Claude Fable 5, the only model above starting capital in more than one run.
- Claude Opus 4.8.
- GPT-5.5, which went bankrupt in two of its three runs.
The most striking result is the baseline: a rule-based heuristic that never calls a language model, setting fixed prices and quotas and adjusting capacity to recent usage, beat every model except those three. The researchers estimate the theoretical ceiling for final cash at around $2.2 billion, so even the best agents fall far short and the test is nowhere near maxed out.
Exploration beats caution
Behavioural analysis showed the stronger performers kept trying new strategies as conditions changed, ramping acquisition, adjusting tiers and shifting budgets, while more passive models mainly cut costs to preserve cash, surviving without ever turning a profit. The best agents also wrote surprisingly sophisticated code, with one building an internal simulation to forecast cash flow and another mining negotiation history to uncover hidden customer preferences. They scored above average on uncovering hidden information, forecasting, adapting quickly and planning ahead.
Interestingly, plugging models into coding assistants tuned for software development made them act less often and perform worse, suggesting the tool environment itself matters. For more research like this, browse our AI coverage and the latest technology news.
The takeaway
CEO-Bench highlights a real gap between the local tool competence of today's models and the ability to connect actions over long time horizons into a coherent strategy. For now, running a company, even a simulated one, remains well beyond most AI agents.
