Looks like I misremember, it's not a study but just a blogpost with some benchmarking. And it's not first, but third: https://news.ycombinator.com/item?id=46582728

I see Clojure actually tops the list! It seems this lines up with what one would've guessed even pre-LLM: the most token-efficient languages are highly expressive languages where concision is also a requirement for idiomicity.

I need to find it, but it was on HN