Reasoning models are pretty interesting for this because they can iterate on the translation to get the subtle parts (like tone) right, especially if you prompt them to do so specifically.