LISP machines, Solo, Edison, Oberon... quite a few systems and languages had that capability if the users and/or developers so desired. In the write-up he gave me, Kay seemed to suggest it had a unique combination of OOP support, conceptual brevity, and especially the late binding. I mean, good performance and stuff too. Those other things were considered key advantages over other systems I named.

Maybe also easier to match to hardware than LISP or FP DSL's along lines of Haskell. I'm a little out of my depth there, though. I just remember LISP machine and OS crowd having to innovate hard to fight performance issues. PreScheme & T being exceptions where low-level was easy.