Automated grading with large language models (LLMs) shows promise on short essays, but thesis-length evaluation remains untested. This study examines alignment between LLM and human grades for master’s theses and the impact of prompt design. English theses from the University of Amsterdam archive are graded on four analytic “Scientific Reasoning” rubrics from the Student Manual Master’s Thesis 2024-2025 (Research Question; Design/Procedure/Methods; Data Analysis; Conclusion/Discussion). Four raters evaluate every thesis: two humans (expert, novice) and two LLMs (ChatGPT-4o, GPT-5). Each LLM uses a task-only prompt and a structured-reasoning prompt. Agreement is quantified with ICC(A,1), quadratic-weighted κ, and Pearson’s r with a Bayesian hierarchical model evaluating effects on absolute residuals. Human resource use is compared to API costs & time use. The study addresses the gap in automated assessment for thesis-length work.