We address the concern that classification of English learners (ELs) primarily based on standardized tests does not accurately reflect what students know and can do with English. While qualitative classroom discourse analyses show that ELs' language-in-use is not as limited as test scores indicate, available evidence is sparse and prone to unintentional biases. Informed by sociocultural theory, we performed a series of quantitative analyses of student conversations during academic tasks at a larger scale. We analyzed 833 conversation transcripts between pairs of K-12 students in science and English language arts (ELA) classrooms. To uncover patterns of lexical resources in different interactions, word frequency and correlation analyses were performed for three dyadic configurations: two English-only students (EO-EO), two ELs (EL-EL), and the heterogeneous combination (EO-EL). Results showed that lexical size was only significantly different among grade-level bands despite EL-EL dyads' conversations being shorter in length in ELA. Relative to EO-EO dyads' language use, in general, EL-EL dyads maintained better word-frequency correlations in science-specific language, while EO-EL dyads maintained strong correlations in ELA-specific language. The findings suggest that attention to authentic language use between peers in class could be more informative than discrete tests and more effective to scaffold EL learning.