Mercor researchers compared twelve licensed CPAs against leading AI models on the APEX Accounting Benchmark, a test suite containing 160 tasks across ten simulated companies. While human accountants with an average of five and a half years experience achieved roughly 37 percent accuracy eighteen months ago, current models like Claude Opus 5.5 now meet 61.8 percent of grading criteria. These systems process structured bookkeeping faster and cheaper than their human counterparts, yet they still fail to close the books without supervision.
The gap persists because the benchmark isolates specific technical skills rather than replicating the full scope of professional practice. The study excluded essential elements such as client communication, internal collaboration, and the application of long-term contextual knowledge. Mercor notes that while AI excels at finding details and following instructions precisely, it cannot yet replicate the judgment required for complex financial oversight. This suggests significant productivity gains are possible, but total replacement of human accountants remains unlikely.
- Claude Opus 5.5 leads the current models with 61.8 percent accuracy.
- Fable 5.1 and GPT-6 Astra follow with 61.0 percent and 57.9 percent respectively.
- The test suite was built by over forty professionals averaging eleven years experience.
