Always use labels for exact indexing--they're more efficient and easier to use, as well as offering things like unique constraints on fields. The legacy indexes are still required for things that require lucene syntax or FTS, but hopefully the new indexes will support that functionality soon.
The main reason the new label-based indexes are more efficient is you pick which nodes you want to index by the label (there's a separate index for each label/property). With autoindexes, there was no way to do this--all nodes were indexed.