This study explored the interplay between content knowledge and reading ability in a large-scale multistage adaptive English for academic purposes (EAP) reading assessment at a range of ability levels across 1–12 graders. The datasets for this study were item-level responses to the reading tests of ACCESS for ELLs Online 2.0. A sample of 10,000 test takers were each time randomly drawn from the test-taking population at five grade clusters without manipulation on proficiency levels, and then with manipulation on proficiency levels. The results indicated that although the bi-factor multidimensional item response theory (MIRT) model fit the data significantly better than the unidimensional two-parameter logistic (2PL) model for Grade 1, no clear evidence can be found regarding the dimensionality of the test for Grades 2–12. However, content knowledge was consistently found to contribute substantially to test performance for low-ability-level test takers across all grade clusters. The findings indicate that EAP reading ability is a multidimensional construct in the onset of EAP reading ability development, but the presence of multidimensionality decreases as proficiency level and grade level increase. This study provides insights into the developmental pattern of the interplay between language and content in EAP reading contexts.