This contribution describes the performance of the ATLAS $b$-tagging algorithms for the 2017-18 data
taking at the LHC. Novel taggers based on soft muons from semi-leptonic decays of the $b$/$c$-hadrons and a Recurrent
Neural Network based on track parameters have been integrated into the final high-level discriminant, based on a boosted decision trees. A new training strategy for the optimization of the multivariate techniques in the high-$p_{\textrm{T}}$ regime will also be presented. Comparisons between data and Monte Carlo simulations and the expected performance for the 2017-18 data taking period will be compared with the former 2016 configuration. The improvements in both, modeling and performance will be discussed.