Abstract Background The correlation and difference in T-cell phenotypes between peripheral blood lymphocytes (PBLs) and the tumor immune microenvironment (TIME) in patients with gastric cancer (GC) is not clear. We aimed to characterize the phenotypes of CD8⁺ T cells in tumor infiltrating lymphocytes (TILs) and PBLs in patients with different outcomes and to establish a useful survival prediction model. Methods Multiplex immunofluorescence staining and flow cytometry were used to detect the expression of inhibitory molecules (IMs) and active markers (AMs) in CD8⁺TILs and PBLs, respectively. The role of these parameters in the 3-year prognosis was assessed by receiver operating characteristic analysis. Then, we divided patients into two TIME clusters (TIME-A/B) and two PBL clusters (PBL-A/B) by unsupervised hierarchical clustering based on the results of multivariate analysis, and used the Kaplan–Meier method to analyze the difference in prognosis between each group. Finally, we constructed and compared three survival prediction models based on Cox regression analysis, and further validated the efficiency and accuracy in the internal and external cohorts. Results The percentage of PD-1⁺CD8⁺TILs, TIM-3⁺CD8⁺TILs, PD-L1⁺CD8⁺TILs, and PD-L1⁺CD8⁺PBLs and the density of PD-L1⁺CD8⁺TILs were independent risk factors, while the percentage of TIM-3⁺CD8⁺PBLs was an independent protective factor. The patients in the TIME-B group showed a worse 3-year overall survival (OS) (HR: 3.256, 95% CI 1.318–8.043, P = 0.006), with a higher density of PD-L1⁺CD8⁺TILs (P < 0.001) and percentage of PD-1⁺CD8⁺TILs (P = 0.017) and PD-L1⁺CD8⁺TILs (P < 0.001) compared to the TIME-A group. The patients in the PBL-B group showed higher positivity for PD-L1⁺CD8⁺PBLs (P = 0.042), LAG-3⁺CD8⁺PBLs (P < 0.001), TIM-3⁺CD8⁺PBLs (P = 0.003), PD-L1⁺CD4⁺PBLs (P = 0.001), and LAG-3⁺CD4⁺PBLs (P < 0.001) and poorer 3-year OS (HR: 0.124, 95% CI 0.017–0.929, P = 0.015) than those in the PBL-A group. In our three survival prediction models, Model 3, which was based on the percentage of TIM-3⁺CD8⁺PBLs, PD-L1⁺CD8⁺TILs and PD-1⁺CD8⁺TILs, showed the best sensitivity (0.950, 0.914), specificity (0.852, 0.857) and accuracy (κ = 0.787, P < 0.001; κ = 0.771, P < 0.001) in the internal and external cohorts, respectively. Conclusion We established a comprehensive and robust survival prediction model based on the T-cell phenotype in the TIME and PBLs for GC prognosis.