Only three AI models end a 500-day startup test with more than their starting capital
Original title:Only three AI models finished above starting capital in a 500-day startup survival test
AI summary
Princeton University introduced the CEO-Bench benchmark, in which AI agents run a simulated subscription software company, NovaMind, for 500 days with $1 million in starting capital. Of the 14 models tested, only Claude Fable 5 (best run profit of $47.15 million), Claude Opus 4.8 ($27.8 million) and GPT-5.5 ($21.3 million) exceeded their starting capital in their best runs. A simple rule-based heuristic that does not call a language model reached $15.76 million through fixed pricing, quotas and targeted development, beating every model except those three. Most models failed to maintain a coherent strategy and went bankrupt before the simulation ended. The test is designed to measure AI's long-term strategic decision-making ability.
Key Takeaways
- Princeton University introduced the CEO-Bench benchmark,in which AI agents run a simulated…
- Of the 14 models tested,only Claude Fable 5 (best run profit of $47.
- 15 million),Claude Opus 4.
- 8 million) and GPT-5.
- 3 million) exceeded their starting capital in their best runs.
Quick Facts
- Published
- Updated
- Source
- The Decoder:AI News(RSS)
- Category
- Papers & Research
- AI score
- 70
普林斯顿的 CEO-Bench 测试了一个反直觉结果,一个不用 AI 的简单规则系统击败了绝大多数模型——在当前 agent 都在比窄任务时,这个测试直指长期战略决策的致命短板,做 agent 的必须看。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Only three AI models end a 500-day startup test with more than their starting capital. June 28, 2026. https://new.nat6.net/en/news/cmqxnrjtc004ssljh9zvejgbo/