... by improving model efficiency so AI tokens are "too cheap to meter" or alternately not that much more expensive to operate than conventional sorts of software?

It's a different view of scaling. I mean a lot of people have this absurd idea of "super intelligence" which is like the cover of this book

https://en.wikipedia.org/wiki/Supermind_(novel)

i mean what does "intelligence quotient 10,000" even mean? On the other hand it's very easy to imagine a world where we have systems like the best ones we have now but we can really afford to use them.