Have you tried comparing runs with HEXINODE and QUADINODE set to ON for both V8.4 and V9? It would be interesting to see if that makes any difference.
It is strange that the error increases with a finer mesh density. In your test models, is there a large variation in stress over a few elements?