Mercor conducted a study comparing 12 licensed CPAs with frontier AI models on accounting tasks. The human participants, who averaged 5.4 years of experience, included senior managers and former Big Four employees. They achieved an average score of approximately 37% while spending between 30 minutes and three hours on the assignments.
In contrast, the Claude Opus 5 model achieved a 100% accuracy rate across 20 attempts, completing each task in under 10 minutes. Mercor noted that the cost per criterion for the AI was $0.21, whereas human accountants incurred costs of $10.35, making the AI approximately 49 times more cost-effective for these specific scenarios.
The test utilized simplified tasks from the APEX-Accounting benchmark, focusing on month-end close procedures such as file searching and data reconciliation. Mercor highlighted that while AI excels at structured, repetitive tasks, it currently lacks the judgment, communication, and contextual understanding required for broader professional accounting workflows.
The study also tracked the evolution of AI capabilities over the past 18 months. While GPT-4o initially scored near zero on these tasks, subsequent models like OpenAI’s o3 and GPT-5 showed significant progress, with recent frontier models reaching or nearing perfect scores on the same benchmark exercises.
