Game-based learning (GBL) may address the unique characteristics of a single subject such as chemistry. Previous systematic reviews on the effects of GBL have yielded contradictory results concerning cognitive and motivational outcomes. This meta-analysis aims to: (a) estimate the overall effect size of GBL in chemistry education on cognitive, motivational, and emotional outcomes compared with non-GBL (i.e., media comparison); (b) examine possible moderators of the effects; and (c) identify the more effective game design and instructional design features (i.e., value-added comparison). We screened 842 articles and included 34 studies. This study is the first GBL meta-analysis that employed a three-level random-effects model for the overall effects. Moderator analysis used a mixed-effects meta-regression model. Results from the media comparison suggest chemistry GBL was more effective for cognition (g = 0.70, k = 30, N = 4155), retention (g = 0.59, k = 20, N = 2860), and motivation (g = 0.35, k = 7, N = 974) than non-GBL and the substantial heterogeneity (I-2 = 86%) for cognitive outcomes. No study reported emotional outcomes, and studies considering value-added comparisons of GBL with versus without specific design features (k = 3) were too few to perform a meta-analysis. Moderator analyses implied that except for publication source and sample size, no other moderator was related to effect sizes. There may be the small-study effects, particularly publication bias. Although we conclude that GBL enhances chemistry learning more than non-GBL, the results also make clear that additional high-quality value-added research is needed to identify design guidelines that may further improve chemistry GBL. More GBL meta-analyses on subjects other than chemistry are also needed. As the first GBL meta-analysis that emphasizes emotion, we call for more research on emotion and on relationships between cognition, motivation, and emotion in GBL.