It really depends on how you selected features, both in your first iteration and your second (as well as how you decided which 10 to remove from the 50 first).

Even though you selected the features based on feature importance, you might not get the most out of them. Some features pairs could be correlated with each other for example, which would mean that adding extra features doesn't give any extra information to your models. Thus, maybe you removed 10 features that were actually interesting, while giving 10 more that were not. Without the specific data, it is hard to tell.