That’s the title of Mercor’s new study: Claude Opus 5 scored 100% when performing four month-end accounting tasks. Licensed CPAs averaged 37%, took up to 18 times longer and cost 49 times more.
Don’t fire your accountant just yet.
The study’s authors spend the most of the paper explaining how narrow their findings were.
The tasks were highly structured and focused on skills models perform well at. The human accountants worked alone, without colleagues to ask or prior company knowledge. Oddly, the “juniors” were licensed CPAs with over five years of experience. It’s hard to see how a scenario where experienced pros only score 37% reflects everyday work.
The colleagues I asked about this mostly shrugged. Some said giving AI to juniors makes things worse, because they become more confident in answers that aren’t quite right.
A junior accountant spends their first couple months following detailed instructions, but they quickly moves on to messier projects and working with customers.
Showing AI can outperform humans on a test is an interesting experiment. But companies don’t care about a test, they care about the whole job.
The detail that the human participants worked alone, without colleagues or prior knowledge of the company, says more about the study than the headline result, because that context is most of what a real month-end close depends on. The observation from your colleagues deserves more attention than the benchmark: if AI makes juniors more confident in answers that are not quite right, the risk moves from speed to review. A business deciding where to use these tools should measure how errors get caught, not only how fast a structured task is completed.