New Delhi. Anthropic’s Claude Opus 5 has outperformed a group of human accountants on a set of structured accounting tasks, completing each assignment in under 10 minutes and achieving a 100% score across all 20 reported attempts, according to a study by AI recruiting company Mercor.
The study compared the AI model with 12 licensed accountants, who had an average of about five and a half years of experience. The human participants were asked to complete four simplified month-end accounting scenarios based on realistic company files. The tasks required them to search through documents, identify relevant figures, perform calculations and submit the results in a specified format.
FCRF Launches CP-FRM to Build India’s Next Generation of Fraud Risk Professionals
The accountants recorded an average score of around 37% on the evaluated criteria. Individual performance varied considerably, with scores ranging from zero to around 90%. Most human attempts took between 30 and 180 minutes to complete. In comparison, Claude Opus 5 achieved 100% across all 20 attempts and completed each task in less than 10 minutes.
The four scenarios covered different accounting situations, including budget variance and cost allocation, hotel revenue and commission reconciliation, event revenue analysis and lease-related calculations. The tasks were designed to include detailed requirements where overlooking a single figure or instruction could significantly affect the final score.
The study also highlighted a substantial difference in the estimated cost of completing the tested work. Using the US median accountant wage as the basis for comparison, Mercor estimated a cost of $10.35 per rubric criterion completed by human accountants, compared with $0.21 for Claude Opus 5. The figures are estimates based on the study’s methodology rather than direct billing costs.
The findings attracted a reaction from technology entrepreneur Elon Musk, who commented on social media that “Super Intelligence (fka AI) is now acing accounting tests.” His remark came after the publication of the Mercor study and focused on the rapid improvement in AI performance on professional tasks.
However, the Mercor researchers cautioned against interpreting the results as evidence that AI can perform the entire job of an accountant. The test measured a limited set of structured tasks that emphasized detailed instruction-following, information retrieval and calculations. Actual accounting work can also involve client communication, collaboration, professional judgment and handling situations that are not represented in the benchmark.
The study also provides a comparison of how AI performance has changed over a relatively short period. Mercor said that around 18 months earlier, leading AI models performed below the average accountant on these tasks. More recent models have since moved beyond the human average, with Claude Opus 5 reaching a perfect score under the study conditions.
The results are therefore focused on specific accounting workflows rather than the profession as a whole. They indicate that AI systems are increasingly capable of handling certain document-heavy, calculation-intensive and tightly defined accounting assignments quickly, while broader professional responsibilities remain outside the scope of the test.
About the author — Suvedita Nath is a science student with a growing interest in cybercrime and digital safety. She writes on online activity, cyber threats, and technology-driven risks. Her work focuses on clarity, accuracy, and public awareness.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
