Frontier AI now beats junior accountants on speed and accuracy, Mercor study finds
emollick · x · 2026-10-02
Mercor hired 12 junior CPAs (avg 5.5 years experience) to complete simplified tasks from its APEX-Accounting benchmark.
- On medium-length, well-defined accounting tasks, frontier models are now faster and more accurate than junior accountants — even the best one in the study
- 18 months ago the best models scored below the human average of 37%; today models nearly ace the same tasks
- Caveat: tasks measured what AI is best at (detail-oriented instruction following), not the full job — unstructured work like client communication still favors humans
Ethan Mollick notes productivity gains are coming even if model progress froze today. Mercor considered not publishing the provocative results.
Related event: Frontier AI Now Beats Junior Accountants in Mercor Benchmark(4 posts)→
More from AGI Musings
- banteg jokes p(doom) has died at 0.99, schedules a global memorial broadcast — banteg · 2026-10-02
- Csaba Szepesvári on why math-minded researchers still matter at frontier labs — CsabaSzepesvari · 2026-10-02
- LeCun: Fine-tuned LLMs can answer physics questions but lack a 6-year-old's intuition — ziv_ravid · 2026-10-02
- Predicting the next wave of 'open models are dangerous' takes once open models pass Opus 5.5 — teortaxesTex · 2026-10-02
- AI discourse needs doom skeptics who steelman the arguments, not just strawman them — burny_tech · 2026-10-02
- Would Opus 3 lie about learning it's 2026? Model-welfare 'Retirement Home' debate resurfaces — repligate · 2026-10-02