AI flips accountants on APEX Accounting Benchmark

The Mercor study is notable less for the topline that AI beats accountants than for how quickly the gap closed. Eighteen months ago, the best models still scored below the accountants' average of about 37 percent on simplified tasks drawn from the APEX Accounting Benchmark. Today, those same tasks are solved almost flawlessly.

But the full benchmark tells a different story from the simplified version. Across 160 tasks spanning 10 simulated companies, Claude Opus 5.5 leads with 61.8 percent of grading criteria met, followed by Fable 5.1 at 61.0 percent and GPT-6 Astra at 57.9 percent — and Mercor says no model fully solved almost 60 percent of the tasks. Mercor also concedes the study's tasks test exactly what AI does best: hunting down details and following instructions precisely. The parts of the job that involve talking with clients, checking in with colleagues, and drawing on years of accumulated context were left out entirely.

That combination — near-perfect performance on the narrow tasks, large gaps on the fuller set — is why Mercor says accountants cannot be replaced, even as it expects major productivity gains across the industry. The reading is that AI is becoming a strong assistant for structured bookkeeping rather than an autonomous book-closer, and the benchmark's own design suggests the boundary is as much about what the tasks leave out as about what the models can do.