Abstract: To analyze the metabolic differences in Wangdu peppers, this study employed high-performance liquid chromatography (HPLC) and gas chromatography-mass spectrometry (GC-MS) to detect chemical components and perform non-targeted metabolomics analysis on five different varieties of Wangdu peppers: Yan Jiao (YJ), La Yan (LY), Guo Ta (GT), Yan Jiao 110 (YJA), and Re La (RL). Machine learning was used to classify and identify the differential metabolites screened. First, HPLC was used to measure the content of capsaicin, dihydrocapsaicin, and Vitamin C (V_C) in the five pepper varieties. Then, GC-MS was used for non-targeted metabolite analysis of the five peppers. Principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) were employed to identify differential metabolites and metabolic pathways. Machine learning methods were used to identify the different pepper varieties based on the differential metabolites. In the five pepper varieties, Re La had the highest content of capsaicin and dihydrocapsaicin, at 533.897±62.187 μg/g and 264.526±28.532 μg/g, respectively. Yan Jiao had the highest V_C content at 146.9±0.029 mg/100 g. OPLS-DA identified 16 differential metabolites, including organic acids such as quinic acid and aconitic acid, which were higher in Yan Jiao, D-sorbitol, which was highest in La Yan, citric acid, D-fructose, D-mannose, and lactic acid, which were most enriched in Guo Ta, D-tagatose and amino acids, which were highest in Yan Jiao 110, and glucose and inositol, which were most abundant in Re La. KEGG pathway enrichment analysis indicated that the differential metabolic pathways mainly included galactose metabolism, fructose and mannose metabolism, the citric acid cycle, starch and sucrose metabolism, glyoxylate and dicarboxylate metabolism, and pyruvate metabolism. Finally, three machine learning methods—random forest (RF), XGBoost, and backpropagation (BP) neural networks were used to classify and validate the differential metabolites of the five pepper varieties. The established classification models achieved accuracies of 100%, 92.9%, and 78.6%, respectively, demonstrating their utility in identifying pepper varieties. These results would provide fundamental data for the quality evaluation, variety improvement, and comprehensive utilization of Wangdu peppers.