Comment by semiquaver

I’ve been wondering this question as well. A thing I’ve noticed much more frequently with opus 5 is straight-up failures to attend to important details even in very recent context. Almost every day I will see it confidently assert very basic and sometimes important things that can be contradicted just be reading a page or two back in the transcript.

This apparent “short-term-memory-regression” is confidence-shattering to me. I don’t feel like I can trust the model to even know things I tell it explicitly. I haven’t seen this behavior to this extent from any model whatsoever, even supposedly much less capable ones, in the year or so I’ve been using them at this extent.