There's a cost calculation even for locally ran models - I guess we should drop that?