For a python library like ours, a 90% coverage threshold should be the ideal. It should already consider the slow tests as well. Coverage is not the only metric that matters, but it is the best one.