That's a useful performance improvement but it's incremental progress in line with what new models often improve over their predecessors, not in line with the much more dramatic reduction they've achieved in training cost.