Claude Opus 5 beats CPAs 100% to 37% on month-end close benchmark, at 2% the cost
McDonaghMatthew · x · 2026-10-14
AI training firm Mercor built a benchmark mimicking a real month-end close, and Anthropic's Claude Opus 5 crushed human accountants on it:
- Accuracy: Opus 5 scored 100% on the rubric; human CPAs averaged just 37%.
- Speed: The AI finished in under 15 minutes, versus 30 minutes to 3 hours for humans.
- Cost: Comparing token costs to the BLS median accountant salary, Opus 5 completed a rubric item correctly for $0.21 versus $10.35 for humans — roughly 50x cheaper.
Mercor's researchers write that "models today already seem more meticulous than junior accountants," which is extremely valuable where small errors are costly. The retweeter adds that consulting revenue overtook audit at half the Big Four (Deloitte, PwC) by 2022, and AI assurance is now outperforming human auditors — putting more pressure on the profession.
More from AGI Musings
- Geoffrey Hinton: 'We're creating alien beings — I'm selfishly choosing humans' — Eastern-Opposite9521 · 2026-10-14
- X user predicts the first AI virus: a small self-replicating model — p_e_cooper · 2026-10-14
- Lenny's Newsletter: what actually counts as a moat in the AI era — HamelHusain · 2026-10-14
- Corporate salary is a fusion of retainer and insurance premium in the AI era — curious_vii · 2026-10-14
- Greg Isenberg: the first AI was a 2021-frozen internet scrape that forgot you at half capacity — building_brocks · 2026-10-14
- Tufa Labs jumps from 56% to 68.91% to reclaim first place on ARC-AGI-3 — burny_tech · 2026-10-14