what if you checked how different models handled trivial changes in SWE benchmark problems?
https://maxxbench.unlap.ai/
quite the interesting data point