The WHO and American Diabetes Association criteria for diagnosing diabetes mellitus assume the presence of a glycaemic threshold with high sensitivity for identifying retinopathy. However, this assumption is based on data from three previous studies that had important limitations in detecting retinopathy. We aimed to provide updated data for the relation between fasting plasma glucose (FPG) and retinopathy, and to assess the diagnostic accuracy of current FPG thresholds in identifying both prevalent and incident retinopathy.We examined the data from three cross-sectional adult populations: those in the Blue Mountains Eye Study (BMES, Australia, n=3162), the Australian Diabetes, Obesity and Lifestyle Study (AusDiab, Australia, n=2182), and the Multi-Ethnic Study of Atherosclerosis (MESA, USA, n=6079). Retinopathy was diagnosed from multiple retinal photographs of each eye, and graded according to the modified Airlie House Classification system. Plasma glucose concentrations were measured from fasting venous blood samples.The overall prevalence of retinopathy was 11.5% in BMES (95% CI 10.4-12.6%), 9.6% in AusDiab (8.4-10.9), and 15.8% in MESA (14.9-16.7). However, we found inconsistent evidence of a uniform glycaemic threshold for prevalent and incident retinopathy, with analyses suggesting a continuous relation. The widely used diabetes FPG cutoff of 7.0 mmol/L or higher had sensitivity less than 40% (range 14.8-39.1) for detecting retinopathy, with specificity between 80.8% and 95.8%. The area under receiver operating characteristic curves for FPG and retinopathy was low and ranged from 0.56 to 0.61.We saw no evidence of a clear and consistent glycaemic threshold for the presence or incidence of retinopathy across different populations. The current FPG cutoff of 7.0 mmol/L used to diagnose diabetes did not accurately identify people with and without retinopathy. These findings suggest that the criteria for diagnosing diabetes could need reassessment.