Epoch's Automation Reports test whether AI can do Epoch's own work

Epoch AI launched Automation Reports to test whether AI models can do Epoch's own research and communication work, and its report says they cannot yet. In the report, Epoch gave six models 11 tasks in five categories (graphic design, data insights, data explorers, AI data center research and research design), with the context a new hire would get.

In its announcement, Epoch says Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra lead but are far from fully automating its work; the report calls them broadly tied. Epoch says leading closed models were reliable on well-defined work such as coding and computational analysis but, on open-ended tasks, often missed implicit style and audience conventions and designed weak experiments.

Epoch lists its own limits: each model ran once per task, a single human grader scored each output, scores are noisy rough indicators, the tasks may not represent other jobs, and differences may reflect each tool's agent harness (the software around the model) as much as the model.