Let me preface by saying that the multiline / inner line tab completion is the main reason I use cursor, as it still leads in actual implementation of this feature, and I’ve been happily using it for close to a year now.
However, in the past month or two I feel the actual quality of the completions has greatly degraded!
This is working in the same exact codebases / languages / types of features that I have been doing all year, and yet now it struggled with even simple completions which it used to one-shot.
One particular failure pattern I have notice is it just directly duplicating lines (sometimes without any changes) or making a very similar line which makes absolutely no sense (ie redefining a const var again immediately after). It seems to completely ignore context around it in favor of mimicking one of the last things you have typed…
Its kinda crazy because even before I loved the way that it would write error messages which actually made sense for the line I’m inserting on. NOw it just duplicates the exact same messages from half way up the function??
Or even the “copy paste” behavior that I liked, it seemed really smart about adding more entries to a list or object or csv or any kind of repetitive task, but once again, now it failing to do that in a smart way, it just dumbly copies now…
This is very sad for me, I’ve actually come to the point today where it was such a hindrance that I had to just turn it off. And then I realized, that it was the only thing keeping me in cursor and not trying some other similar IDE…
I’m not sure if this is like due to changed prompting / the way context is presented to the model / a smaller/faster (cheaper) model / some kind of collapse as the model is progressively trained… But it’s really unfortunate, given how strong it was only 4-5 months ago. Like before it really was in tune with what I was thinking while coding, it often correctly predicted EXACTLY what I was about to type. But I have not experienced that in a while (again working on basically the exact same codebase / features / styles / etc)
I would happily have a slower model which only provided completions infrequently, but with quality completions. Or a way to go back to the older model. Maybe a hotkey that triggers the smarter completion, rather than just doing it on every character I type.
Also I don’t think it’s just me getting used to it and then applying selection bias to the failure cases, like I said I used to have exactly what I was about to type happen MULTIPLE times a day, but it has not happened in weeks now, thats a cold hard fact. SOMETHING HAS CHANGED
I know tab doesn’t get as much love as the “vibe coding” composer, but I really do like to write most of my own code, and before tab was a speed multiplier for that. I don’t feel that way any more 
Please don’t take this as just being negative, I just really am sad that the performance has degraded and wish that it had not happened to such a great product.
