Imo deepseek 4.1 (and a lot of the cheaper models, Luna is similar) show the issues in benchmarks. At this point.

In actual day to day development the differences are a lot harder to spot. Maybe deepseek is worse, but I asked it to run until it was able to launch itself and verify it worked as expected, and it did. Maybe it wasted some turns, idk, but when it said it was done, it was done.

I have no doubt there's things it's worse at, but what percentage of development is truly novel?