For a long time, large language models (LLMs), known for their expertise with language and reasoning, miserably failed at serious finance tests. But now all that is changing; a new study has found that six leading AI models passed all three levels of Chartered Financial Analyst (CFA) certification exams. This is often reckoned as the ‘Gold Standard’ for investment professionals, opening doors for global opportunities and higher rewards. Perhaps this is why it is widely considered to be one of the most difficult exams to crack.
The study revealed that Google’s most capable model, Gemini 3.0 Pro, secured a record high score of 97.6 per cent on level 1. The researchers tested Gemini 3.0 Pro and GPT-5. Claude Opus 4.1, Grok 4, and DeepSeek-V3.1 on 980 questions across all exam tiers. While GPT-5 topped Level II with 94.3 per cent, Gemini 3.0 Pro performed well in the most difficult constructed-response section with 92 per cent.
The research shows that AI has largely overcome earlier limitations with CFA-style problem-solving. The latest reasoning models are not only passing the CFA exams, but most of them are clearing all three levels with ease.
Why is this a big deal?
The Chartered Financial Analyst (CFA) programme is among the toughest professional qualifications in finance. It has three levels, each designed to test a different kind of thinking capability. Level I is about foundations, and it features multiple-choice questions checking whether a candidate understands the core concepts, such as economics, financial reporting, ethics, and quantitative methods.
Level II is about application, and here questions are grouped into case studies. Candidates are given briefs and are expected to apply relevant concepts correctly and not simply recall formulas. However, Level III is the real mammoth, as it tests a candidate’s synthesis and judgement. Here, the candidates are expected to build portfolios, evaluate strategies, and explain their reasoning in writing.
Finance is complicated, as it involves maths, judgement, ethics, and context. And this was the main reason why earlier models struggled, as they were mostly able to recall facts but often missed nuance or wrongly applied rules.
What did the AI models do differently?
As part of the research, the AI models were tested on 980 mock CFA questions across all three levels. The questions were actually drawn from current CFA Institute practice materials and reputable mock exams that are based on the latest curriculum.
The researchers tested older models like ChatGPT and GPT-4 to recreate earlier results and later compared them with newer reasoning-focused models such as GPT-5, Gemini 3.0 Pro, Grok 4, Claude Opus 4.1, and DeepSeek-V3.1. They also employed a strict pass criterion similar to the real exams. Almost all the modern reasoning models passed Level I, Level II, and Level III. It needs to be noted that the result is a dramatic jump from just a couple of years ago, when even GPT-4 struggled to get through Level II.
What has changed in newer AI models?
According to the researchers, the key difference is architecture, as the newer models are better at multi-step reasoning, and they do not just retrieve answers. The models can connect ideas, track context across long case studies, and apply rules based on the conditions.
In simple words, the researchers explained that the older models often failed at ethics questions, and this was not because they did not know the rules, but because they applied them mechanically. In contrast, the newer AI models, while they still make mistakes, do so far fewer of them, and often for subtler reasons.
Another major advancement in new models is their ability for quantitative reasoning. While earlier research found that math-heavy sections were a weakness. In this study, top models show near-zero error rates in many quantitative topics.
The models still stumble, as the study shows that ethics remains a weak spot, as even the best models tend to make mistakes in ethical and professional standards. Their performance in the constructed-response section also raises concerns. The researchers used another AI model to assess the written responses, which introduces bias, such as longer, well-structured answers scoring better even if they contain subtle errors. However, human CFA graders would be much more critical.
Can AI replace CFA analysts?
No, this doesn’t mean AI can replace CFA charterholders tomorrow. Passing a competitive exam does not equate to managing money, engaging with clients, or navigating the volatile market scenarios. However, it does mean something important – AI has crossed a threshold in professional reasoning. These models are now capable of handling structured, high-stakes financial reasoning tasks that were previously considered out of reach.
But, for finance professionals, this is both useful and troublesome. AI can increasingly act as a junior analyst who seldom rests or complains and sometimes gets ethics wrong. However, for regulators and educators, it raises bigger questions – if AI can pass tough professional exams, the value of those exams shifts from testing knowledge to testing judgement, accountability, and real-world decision-making.
The latest reasoning models are not only passing the CFA exams, but most of them are clearing all three levels with ease. (Express Image/FreePik) 
Share
