Approximate Bayesian computation (ABC) is a powerful tool for model-based inference of demographic histories from large genetic data sets. For most organisms, its implementation has been hampered by the lack of sufficient genetic data. Genotyping-by-sequencing (GBS) provides cheap genome-scale data to fill this gap, but its potential has not fully been exploited. Here, we explored power, precision and biases of a coalescent-based ABC approach where GBS data were modelled with either a population mutation parameter (θ) or a fixed site (FS) approach, allowing single or several segregating sites per locus. With simulated data ranging from 500 to 50 000 loci, a variety of demographic models could be reliably inferred across a range of timescales and migration scenarios. Posterior estimates were informative with 1000 loci for migration and split time in simple population divergence models. In more complex models, posterior distributions were wide and almost reverted to the uninformative prior even with 50 000 loci. ABC parameter estimates, however, were generally more accurate than an alternative composite-likelihood method. Bottleneck scenarios proved particularly difficult, and only recent bottlenecks without recovery could be reliably detected and dated. Notably, minor-allele-frequency filters - usual practice for GBS data - negatively affected nearly all estimates. With this in mind, we used a combination of FS and θ approaches on empirical GBS data generated from the Atlantic walrus (Odobenus rosmarus rosmarus), collectively providing support for a population split before the last glacial maximum followed by asymmetrical migration and a high Arctic bottleneck. Overall, this study evaluates the potential and limitations of GBS data in an ABC-coalescence framework and proposes a best-practice approach.