We performed a meta-analysis of approximate number system (ANS) training studies to investigate the strength of the causal effects of practicing ANS related tasks on symbolic math performance. Across 33 effect sizes from 11 studies involving 754 participants, for which neither the treatment nor control group received symbolic training, we found a small nonsignificant and sensitive effect of ANS training on symbolic math task performance (g = 0.11, 95% confidence interval, CI [-0.01, 0.22]; precision-effect estimate with standard errors (PEESE) adjusted g = -0.04, 95% CI [-0.58, 0.50]). Some heterogeneity was accounted for by participant age, with larger estimates for adults than for children. Estimates did not vary significantly by ANS training type, training duration, and control group type. An exploratory analysis on the transfer effects of ANS training on untrained nonsymbolic tasks suggests weak support for the key auxiliary assumption that ANS training has substantial effects on a general ANS, indicating that the training literature may not adequately represent theories of how ANS influences symbolic number performance.