The Pulse
Mercor Finds Claude Outperforms CPAs on Four Close Tasks
Mercor says Claude Opus 5 scored perfectly on 20 attempts at four adapted month-end accounting tasks, outperforming 12 licensed CPAs in the study. The company cautions that the test measured structured file-search and instruction-following
AI.info Team · Written with AI ·
“On medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants.”
Aden Barton, Mercor
Mercor published that finding on October 1, 2026, after comparing 12 licensed accountants with AI models on four adapted month-end close tasks. Its paper says Claude Opus 5 completed every one of 20 attempts with a perfect score, while the accountants’ unassisted attempts averaged 37% of the rubric criteria. The result applies to a narrow set of structured tasks, not to accounting work as a whole.
Four close scenarios, one perfect AI score
The accountants worked through simulated company files to find figures, apply the company’s accounting rules, perform calculations and deliver summary tables. The scenarios covered nonprofit budget allocations, hotel revenue and commissions, event revenue analysis, and lease schedules. Mercor adapted the tasks from its APEX-Accounting benchmark, simplified them for human completion and supplied a file map to reduce search time.
Each task had a three-hour limit for participants. In the unassisted comparison, the 12 accountants made 23 task attempts; Claude Opus 5 made 20. Mercor reports that the human attempts took roughly 30 to 180 minutes, while Claude finished each attempt in under 10 minutes. The company says its graders scored work against accountant-written rubrics; eight human submissions were also hand-graded, with 98% agreement between those grades and the language-model judge.
The “junior accountant” label needs context
Mercor’s headline and summary describe the participants as junior accountants, but the paper’s appendix shows a broader mix. All 12 held active U.S. CPA licenses and averaged 5.4 years of accounting experience. Seven described their current level as senior accountant or senior auditor, three as manager or above, and two as staff or associate; none were partners, directors or executives.
That detail matters when interpreting the comparison. The study’s average human score was 37%, but Mercor’s task authors had expected about 30% for a junior and 55% for a mid-level accountant. The authors say difficult-to-spot requirements could compound: missing one figure or instruction could pull down the score even when much of a submission was correct.
AI assistance did not raise the score
The experiment also tested accountants using Claude Cowork with Opus 5 on two of their four tasks, with task order and AI-assisted assignments randomized. Mercor reports that Claude working alone was slightly more accurate and about 15 times faster per task than accountants working with Claude. In three of the four assisted attempts that fell short of a perfect score, the model had produced the correct figure at some point, but it did not survive into the final submission.
That finding does not establish that human review is generally harmful. Mercor designed the study to measure how much AI helped accountants, but the model’s perfect scores left no room for improvement under the rubric. The setup also gave participants no coworkers to consult and no accumulated knowledge of the fictional companies—context accountants commonly build on the job.
A measure of file work, not the whole profession
Mercor says the tasks favored capabilities models handle well: searching across files, following detailed instructions and catching small requirements. They did not test client communication, clarification, collaboration or the tacit knowledge that develops over time. The paper notes that five participants rated the setting unrealistic, mainly because it lacked interaction with colleagues; seven of the 12 rated the tasks realistic.
The cost comparison is similarly specific. Using Claude Opus 5’s listed token prices and the U.S. median accountant wage, Mercor estimates $0.21 per rubric criterion met for the model and $10.35 for accountants—49 times more for the human work. The model figure uses token prices without caching discounts, while the wage figure is not what participants were paid. Those assumptions make the study a comparison of task performance and estimated cost, not a general calculation of what it costs to replace an accountant.