AI Vs CFA Level III: 23 Models Tested, Top Score 79.1%—See Who Leads

Sunita Somvanshi

Photo Source :- Google

Interactive ranks: AI on mock CFA Level III

Built from your supplied text and verified sources. Concise facts, interactive ranks, and neutral recap.

What’s new: A research collaboration between NYU Stern’s Professor Srikanth Jagabathula and Shilpi Nayak, Co-Founder & CTO of Goodfin, tested 23 AI models on mock CFA Level III materials and reported composite scores above the typical passing threshold for several frontier systems. Methods and results are posted on the CFA Benchmark site, the preprint, and in the research announcement.

Key Findings

  • Top performer: OpenAI’s o4-mini achieved a 79.1% composite score
  • Gemini 2.5 Flash reached 77.3%, while Claude Opus 4 scored approximately 74.9%
  • For context, the CFA Institute reported a 43% Level I pass rate for August 2025
  • Human candidates typically invest around 300 hours of study per level
Student writing answers on an exam paper at a classroom desk during a test.
A student focuses on answering exam questions. Will exams become obsolete in future?
23
AI models evaluated
79.1%
Top composite (o4-mini)
77.3%
Gemini 2.5 Flash composite
43%
Level I pass rate (Aug 2025)

Leaderboard (user-supplied list; verified notes added)

Rank Provider Model Overall MCQ Essay Notes
Verified figure User-supplied entry Reasoning Non-reasoning

Data notes: “Verified” badges reflect values corroborated by the preprint and/or CFA Benchmark (e.g., o4-mini 79.1%, Gemini 2.5 Flash 77.3%, Claude Opus 4 ~74.9%). Remaining rows are from the supplied list.

Performance Highlights

The leaderboard shows reasoning-enabled models generally outperformed non-reasoning models, with particular strength in essay responses. Top systems demonstrated competency in portfolio management and wealth planning concepts.

Implications

While these results are promising, they represent performance on mock exams only. Actual CFA certification requires program enrollment, fees, and approximately 4,000 hours of relevant work experience. Organizations implementing AI in financial analysis should maintain human oversight, review processes, and clear audit trails.

See our infrastructure coverage: custom AI chips, NVIDIA Blackwell & MLPerf, GB200 NVL72 throughput, compute deals, Gemini 2.5 analysis, and GPT-4.5 overview.

The section presented the user-supplied ranks, added verified notes for select entries, summarized the 23-model mock evaluation, and referenced the official CFA pass-rate page and benchmark sources. Links to internal explainers and infrastructure coverage were included for reference.

Leave a comment