Claude Opus 5 outperforms certified accountants in accounting tasks

Summary

According to a study shared by Rohan Paul on the X platform, Mercor found that Claude Opus 5 achieved a perfect 20/20 score in minutes on medium-length, well-defined accounting tasks. In contrast, 12 licensed CPAs showed variable completion rates, performing between 0% and 90% within a 3-hour timeframe. Based on the evaluation criteria, the AI's cost was $0.21, while the human accountant's cost was measured at $10.35.

Why it matters

If frontier models are now competing with or outperforming junior accountants in clearly defined, rule-based tasks, the economic case for automating broad segments of back-office finance is no longer hypothetical. The 18-month trajectory from underperforming to outperforming human benchmarks indicates that this gap is widening, forcing companies to rethink headcount and workflow design. This also signals that 'well-defined' professional tasks, once considered safe human domains, are becoming the areas most rapidly displaced by automation.

Highlights

• Claude Opus 5: 100% accuracy, minutes per task, $0.21 per criterion. CPAs: 0–90% accuracy, up to 3 hours, $10.35 per criterion
• Ethan Mollick: Frontier AI is faster and more accurate than junior accountants on medium-length, well-defined tasks
• 18 months ago, comparable models scored significantly below human accountants in similar benchmarks

Critical take

The evidence base relies on a single, non-peer-reviewed Mercor blog post testing only 12 CPAs on *well-defined, medium-length* tasks. While Mollick carefully specifies 'junior accountants,' the headline claiming to 'outperform certified accountants' overstates the scope. A perfect score on 20 scenarios says little about the value licensed professionals still add in ambiguous, multi-stakeholder, or judgment-heavy accounting work.