Various similar primary studies report varying impacts of applying blended problem-based learning (Blended PBL) in improving students' 21st-century skills. This study aims to group similar studies on the effect of blended PBL on students' 21st-century skills in science learning through meta-analysis. In addition, studies were analyzed on variations in models, education levels, subjects, and 4C skills. There are 44 articles worth analyzing. The data was analyzed using Excel and JASP applications by presenting results in size effect values, forest plots, and published bias tests. The results of the analysis show that science learning with blended PBL improves 21st-century skills that are better at science learning than other learning (g = 0.617, 95% Cl 0.296-0.937, p < 0.001, I[superscript 2] = 96.861%). Analysis of moderator variables showed that in the variation model, level of education, subjects, and 4C skills, there was a significant effect. These findings confirm that blended PBL is efficiently used to train students' 21st-century skills in science learning.