Infants with MLL-rearranged infant acute lymphoblastic leukemia (MLL-r iALL) undergo intense therapy to counter a highly aggressive malignancy with survival rates of only 30–40%. The majority of patients initially show therapy response, but in two-thirds of cases the leukemia returns, typically during treatment. The glucocorticoid drug prednisone is established as a major player in the treatment of leukemia and the in vivo response to prednisone monotreatment is currently the best indicator of risk for MLL-r iALL. We used two different single-cell RNA sequencing technologies to analyze the expression of a prednisone-dependent signature, derived from an independent study, in diagnostic bone marrow and peripheral blood biopsies. This allowed us to classify individual leukemic cells as either resistant or sensitive to treatment and show that quantification of these two groups can be used to better predict the occurrence of future relapse in individual patients. This work also sheds light on the nature of the therapy-resistant subpopulation of relapse-initiating cells. Leukemic cells associated with high relapse risk are characterized by basal activation of glucocorticoid response, smaller size, and a quiescent gene expression program with cell stemness properties. These results improve current risk stratification and elucidate leukemic therapy-resistant subpopulations at diagnosis.