Noticed the same. Also, I noticed the same with the prev version, immediately improved when switching to the new 5.2.

The smoking gun is the time. If I ask it a question that's subtle in "thinking" mode and it starts replying in a few seconds, the answer will probably be trash. I'm almost sure they degrade the models over time.

An interesting approach - start strong when everyone is running benchmarks, taper off through the life of the release, introduce new version that is no better than the last version but it magically seems much better by comparison to the degraded previous version. Rinse, repeat, grift.