Can AI automate Epoch?

We gave six models real work tasks from Epoch, such as graphic generation and research design, to test how close they are to automating our work. While frontier models such as Claude Fable 5.1 and GPT-6 Astra perform reliably on well-defined tasks, they fail in the more open-ended aspects that prevent full automation. Epoch Automation Reports captures these shortcomings. We curate a diverse set of tasks directly from our work at Epoch, provide the models with the necessary context, use the highest available reasoning settings, and have them attempt the tasks without further intervention. We then manually grade the outputs against our own employee standards. These evaluations aim to capture gaps in typical benchmarks, which concentrate on easily verifiable tasks that capture only a subset of real work. These existing benchmarks may also overstate progress if a model has been optimized to score well on them. By combining the systematic approach of benchmarks with the holistic impressions that come from testing models on real work, we expose capability gaps that existing benchmarks miss and inform broader questions about AI adoption. Although Fable 5.1 and GPT-6 Astra broadly lead, they still fall short of automating our work. Open-weight models lag further…