The h2o API doesn't send the type of estimator after returning a new model (in JSON response). The type of estimator is assigned to model in R and Python client after receiving a response. It happens in model train method and grid search method separately. This cause inconsistency and can cause problems in the future if a new unsupervised method will be added (for example isolation forest is not on the list now).
The idea is to return a type of estimator in JSON response and set it in the client as the property of a model.
This improvement is associated with the bug (GIT PL with the discussion).