> 90+% of queries are probably so common/evergreen that, if a cheaper model made all the different language vari[a]nts and ways of asking the question into one query, they could be cached quite effectively.

As I wrote in some parallel post:

"When I look at my history with AI chatbots, 90 % of the conversations revolve around research-level problems in math and computer science. :-D"

For quite a few of them, creating a decent answer involved quite a bit of thinking by the AI, so I assume they are not the easiest queries to answer, and the queries are too obscure to cache.