Johnson's algorithm appears to run Bellman–Ford as a sub-routine.

At that point you are already spending |V||E| time, so you are not getting anywhere close to the performance of Dijkstra and the new algorithm.

But I guess you are saying that this preprocessing is OK if you need to run many single-source-shortest-path queries on the graph.