I thought I was going crazy. I was late to upgrade from 4.8 to 5 and I already want to go back. I’ve noticed that in addition to taking liberties with my instructions, it is also less capable at debugging its own issues. In trying to fix a problem in my CI pipeline, it went chasing some “quadratic race condition” it claimed. It turned out that it had added a loop in a test at the wrong level and it spent an hour chasing it down before I realized what was going on. I pointed 4.8 at the same problem and it solved it in 2 minutes.

I certainly agree with the original post. It feels like the model has been highly benchmark tailored and it is now worse at solving problems that fall outside of the standard patterns.