I suggest reading the Mythos report's discussion on SWE-bench and contamination. I think it's fairly convincing that you can account for contamination and still trust SWE-bench numbers on models that aren't over-optimized for it.