Mercor, a company that builds evaluation benchmarks by testing AI against real human professionals, published a study this week comparing frontier AI models against 12 licensed, practicing CPAs on real month-end close accounting work — the file navigation, reconciliation, and detail-checking tasks that make up a real chunk of junior accounting labor. The results are stark enough that they're worth walking through carefully, caveats included, rather than reducing to a single headline number.
Claude Opus 5 scored 100% across all 20 attempts it was given. The human accountants' accuracy ranged from 0% to about 90% on the same tasks. That's not a close contest on the specific, narrow slice of work the study measured — and the cost difference behind those numbers is just as wide.
Quick summary: An October 2026 study from Mercor tested frontier AI models against 12 licensed CPAs on simplified, real month-end close accounting tasks. Claude Opus 5 scored 100% across 20 attempts, each completed in under 10 minutes; human accountant accuracy ranged from 0% to roughly 90%, with most attempts taking 30 to 180 minutes. Cost per rubric criterion met: roughly $0.21 for the AI (an upper bound, per Mercor) versus $10.35 for a human at the US median accountant wage — about 49 times cheaper. Important limits: the tasks tested were detail-oriented and well-defined (file navigation, reconciliation), not client communication or the messier judgment calls that come up mid-close. On Mercor's harder APEX-Accounting benchmark, no model completed every task, and without human oversight, AI still can't close a company's books on its own.
The Numbers That Matter
Claude Opus 5 completed all 20 of its attempts correctly, each one in under 10 minutes. The 12 licensed CPAs in the study ranged from 0% to roughly 90% accuracy on the identical tasks, with most attempts taking between 30 and 180 minutes — three to eighteen times longer than the AI, even before accounting for the accuracy gap (Source: the-decoder).
The cost comparison is where this becomes an economic story rather than just a capability one: roughly $0.21 per rubric criterion met for the AI versus $10.35 for a human accountant at the US median wage — the AI runs at roughly a 49th of the cost, on top of finishing faster and more accurately on this specific task set (Source: Mercor).
| Claude Opus 5 | Licensed Human CPAs | |
|---|---|---|
| Accuracy | 100% (20/20) | 0%–90%, highly variable |
| Time per task | Under 10 minutes | 30–180 minutes |
| Cost per rubric criterion met | ~$0.21 | ~$10.35 |
Before You Panic: What the Study Didn't Test
The tasks measured here were deliberately narrow: detail-oriented, well-defined, instruction-following work like file navigation and account reconciliation. The study did not test client communication, judgment calls that require weighing ambiguous or incomplete information, or the messier parts of a real month-end close that rarely show up in a clean benchmark. On Mercor's own harder APEX-Accounting benchmark, no model — including Claude Opus 5 — completed every task successfully, and the researchers' own conclusion is explicit: without human oversight, AI still can't close a company's books on its own (Source: the-decoder).
What This Means If You Work in Accounting
The honest read is that the specific, repetitive, well-defined slice of junior accounting work — the part this study measured — is now something frontier AI handles faster, more accurately, and at a fraction of the cost. That's a meaningfully different claim than "AI can do an accountant's job," since the judgment-heavy and client-facing parts of the role weren't tested and, per Mercor's own harder benchmark, still aren't something any model handles reliably without a human checking the work. LumiChats has covered the broader data on AI's impact on entry-level roles elsewhere — this study is a concrete, task-level data point inside that larger pattern, not a verdict on the profession as a whole.
If you're early career in accounting, the practical move isn't to assume the job disappears — it's to get good at the parts this study didn't test: client communication, ambiguous judgment calls, and reviewing AI-generated work for the errors a model won't flag on its own.
01Did AI really beat human accountants, or is this exaggerated?
On the specific tasks tested — detail-oriented, well-defined month-end close work like file navigation and reconciliation — yes, Claude Opus 5 scored 100% against a 0%–90% range for 12 licensed CPAs, at a fraction of the cost and time. The study did not test client communication or harder judgment calls, where AI still falls short.
02Can AI fully replace accountants now?
No, according to the same study. On Mercor's harder APEX-Accounting benchmark, no model completed every task, and the researchers explicitly concluded AI still can't close a company's books without human oversight.
03How much cheaper is AI at these tasks?
Roughly 49 times cheaper — about $0.21 per rubric criterion met for the AI versus $10.35 for a human accountant in this study — while also finishing in under 10 minutes versus 30 to 180 minutes for the human attempts.
04Which AI model was tested?
Claude Opus 5 is the model whose 100% accuracy score is highlighted in reporting on the study; Mercor's research also evaluated other frontier models, though Opus 5's results are the headline figure circulating.
If a study like this makes you want to see how different AI models actually handle detail-heavy work you care about — not just a benchmark's narrow task set — LumiChats gives you one subscription with 40+ AI models, including Claude Opus 5, so you can test your own real tasks against several models before trusting any single benchmark's verdict.
