40% vs 90%? Sure.
70% vs 90%? _Absolutely meaningless_ as you are not measuring coding intelligence but “how well can the model cheat flaws in SWEBench Verified”, the former can certainly be better at coding even assuming no deliberate benchmaxxing / foul play.