Abstract It has been hypothesized that incidence of and mortality from several malignancies are increased among diabetic patients. Whether certain treatment modalities, including use of metformin, sulfonylureas, or insulins, affect cancer incidence or mortality and whether use of long-acting insulin analogues glargine and detemir may increase cancer incidence more than traditional human insulins are debated. The objective was to investigate the association between specific glucose-lowering agents and cancer incidence in diabetic members of an Israeli health maintenance organization. We studied a cohort of 36 342 diabetic patients aged at least 18 years with no history of cancer or treatment with insulin as of January 1, 2003. For the period from January 2003 to December 2007, we searched pharmacy records for purchases of glucose-lowering agents, including metformin, sulfonylureas, human insulin, and analogue insulins. Incident cancer diagnoses were identified from the health maintenance organization cancer registry. We studied the association of cancer incidence with the use of specific glucose-lowering agents, controlling for age, sex, and baseline glycohemoglobin measurement. Cancer was diagnosed in 6% of the study cohort during 164 652 person-years of follow-up time. Cancer incidence increased with age and varied with medication purchasing patterns. On multivariate analysis, age (hazard ratio [HR], 1.049; confidence interval [CI], 1.045-1.052), male sex (HR, 1.16; CI, 1.065-1.264), and number of insulin purchases (HR, 1.007; CI, 1.001-1.012) were significantly associated with increased cancer risk, whereas number of metformin purchases was associated with reduced cancer risk (HR, 0.996; CI, 0.994-0.998). Male sex, age, and human insulin purchases were associated with increased cancer incidence, whereas metformin purchases were associated with decreased cancer risk. There was a trend for increased cancer incidence associated with use of long-acting insulin analogues, but the number of long-acting insulin analogue users was too small for risk estimates to be conclusive.