The report says six models attempted 11 tasks across five categories, including graphic design, data insight generation, data explorer creation, AI data center research, and research design. Epoch provided task context and relevant tools, then manually graded outputs against its employee standards.
The report finds that Fable 5.1 and GPT-6 Astra broadly lead, and frontier models perform reliably on well-defined tasks, but the tested systems do not yet autonomously produce Epoch-quality work. The authors identify difficulties in learning implicit standards, designing experiments that measure their claims, and distinguishing flawed results from key findings. Open-weight models lag further behind, including on some more clearly specified work. The report describes an 11-task evaluation suite; it does not establish that every occupation or organization would see the same results.