The specifics:
Researchers used 980 questions from all exam levels to test GPT-5, Gemini 3.0 Pro, Claude Opus 4.1, Grok 4, and DeepSeek-V3.1.
Gemini 3.0 Pro dominated the most challenging constructed-response portion with 92%, while GPT-5 led Level II with 94.3%.
In 2023, GPT-4 only passed Level I and GPT-3.5 failed the first two levels; it took around 24 months to reach near-perfect ratings.
In September, a study conducted at NYU also revealed models that passed all three criteria, although their scores were in the 70s rather than the near-perfect numbers of existing frontier systems.
Gemini 3.0 Pro dominated the most challenging constructed-response portion with 92%, while GPT-5 led Level II with 94.3%.
In 2023, GPT-4 only passed Level I and GPT-3.5 failed the first two levels; it took around 24 months to reach near-perfect ratings.
In September, a study conducted at NYU also revealed models that passed all three criteria, although their scores were in the 70s rather than the near-perfect numbers of existing frontier systems.
Although passing a standardized test and managing the day-to-day demands of financial analysis are still very different, the rate of improvement on these tests is rapid, and models who become proficient in finance could shift the profession's value toward human skills like relationship management and client judgment.