AI can now pass the hardest level of the CFA exam in a matter of minutes (www.cnbc.com)

Researchers from NYU Stern and AI wealth-management startup GoodFin evaluated 23 large language models on mock CFA Level III exams and found several frontier “reasoning” models — notably o4‑mini, Gemini 2.5 Pro, and Claude Opus — can pass the previously intractable final test in minutes. The breakthrough hinges on chain-of-thought prompting, which lets models generate stepwise reasoning to handle the Level III’s essay-style portfolio-management and wealth-planning problems that earlier work (two years ago) showed were hard for AI. For context, humans typically invest roughly 1,000 study hours across years to earn the CFA, so the speed and depth of current models mark a big leap. The result matters because it demonstrates that LLMs are approaching competency for specialized, high‑stakes analytical tasks, suggesting immediate productivity gains in financial research, portfolio construction, and client-facing drafting. Authors and GoodFin’s CEO Anna Joo Fee expect the technology to transform workflows but not fully replace human CFAs — machines still struggle with context, intent and nonverbal cues. Key implications: chain‑of‑thought and advanced model architectures enable complex written reasoning, but deployment will require careful oversight, calibration to avoid hallucinations, and regulatory and ethical guardrails before these systems are used in real-world fiduciary or licensing contexts.