Mercor Study: Claude Opus 5 Outperforms Certified Accountants on APEX-Accounting
Mercor published a human-baseline study for its APEX-Accounting benchmark, hiring 12 CPAs (avg 5.5 years experience) to complete 4 simplified real-world month-end close tasks and comparing them against generations of AI models. As of May 2025 humans still beat AI overall, but today's frontier models are faster and more accurate than junior accountants — and even the best-performing accountant in the study. Elon Musk's repost amplified the discussion.
Confirmed
- The study by Mercor proposes establishing human baselines for benchmarks, using 4 real accounting tasks
- The 12 CPAs averaged 37%, with individual scores ranging from 0% to 90%; several failed to finish within the 3-hour limit
- Claude Opus 5 scored 100% on all 20 attempts, under 10 minutes per task; GPT-4o scored near zero in mid-2024, and AI crossed the human average sometime in 2025 — a leap from near-zero to perfect in about two years
- @emollick relayed the core finding; @paulcal noted the best human accountant was also surpassed; @rohanpaulai and @markk supplied score details
- @elonmusk reposted, saying superintelligence (formerly AI) can now ace accounting tests, noting that just 18 months earlier the best model fell short of the 37% human average; researchers called the results startling
Why it matters
- The result offers a rare quantified reference for how fast AI may displace junior white-collar work. @pswider noted the 18-month catch-up speed and sketched the industry rhythm: medical AI two years ago, education AI a year ago, accounting now
- @rohanpaulai argued rule-based work is no longer a human moat, citing the jump from near-zero to perfect within about two years
- @paulcal quipped that the attitude toward Claude is shifting from the derogatory "did you use AI?" to the obligatory "did you use Claude?", reflecting normalization of AI in professional work
- Reddit and other platforms focused on the impact on finance jobs and the methodology of benchmarking against human capability, showing high interest in job-displacement debates
2026-10-02 ~ 2026-10-03 · 11 related posts
Primary sources
- Frontier AI models now beat CPAs on accounting tasks that humans won as of May 2025 — soumitrashukla9 · 2026-10-02
- [source] Frontier AI now beats junior accountants on speed and accuracy, Mercor study finds — emollick · 2026-10-02
- Frontier AI now beats junior accountants on defined tasks, study finds — pswider · 2026-10-02
- Human baselines for benchmarks: AI now outperforms junior accountants, Mercor says — ResultBackground2450 · 2026-10-02
- Study: Frontier AI beats even the best accountants on detail-oriented tasks — paul_cal · 2026-10-02
- [source] Musk: 'Super Intelligence' Now Aces Accounting Tests AI Failed 18 Months Ago — elonmusk · 2026-10-02
- AI scores 100% on accounting tasks, up from near zero two years ago and far past the 37% human average — rohanpaul_ai · 2026-10-02
- Claude Opus 5 scores 20/20 on CPA tasks while 12 licensed accountants top out near 90% — rohanpaul_ai · 2026-10-02
- [source] Claude Opus 5 scores 100% on accounting tasks where licensed CPAs average 37% — mark_k · 2026-10-03
2 near-duplicate retellings: rohanpaul_ai · MFordFuture