I miss the old Opus 4.6 too. They're probably quantizing the old models.
K/V cache compression and context shortening / summarisation. And yes, I suspected Quants too.