I’ve been using 0.45.17 and it’s just so good. I don’t know why. Is it the prompts?
Maybe it sticks more to the letter of the law due to smaller context window so attention on the requested task is stronger?
I’ve tried 0.46 and most recently 0.47, with MAX.
I was able to quickly bang out a very complex session management improvement to our Next.js codebase with 0.45 that I could not get 0.47 to do after a much longer conversation and far more elaborate attempts.
I got a very satisfactory result after just 2 correction requests (one of which was my own fault due to insufficient upfront definition) in 0.45.
Where as MAX / 0.47 (even in just Ask mode, which seems to be closest to Chat) repeatedly kept trying to implement these more complex systems (needlessly ignoring / recreating existing systems from scratch, and generally doing too much), and was just consistently wrong in its interpretations of my asks.
At one point it create this one else-if case and showed that it completely misunderstood my ask (related to long term and short term session management).
Meanwhile 0.45 just knocks it out of the park.
So confusing. Almost feels like it needs a new whole number for the jump to 0.46 or something.
I didn’t notice such a stark breakdown in functionality & productivity before 0.45.
I have not got one of the latest models to fix a single bug / build any new functionality successfully (despite on average using them for ~5x as long as 0.45 per task)
I think 0.45 is excellent (for me even better with o3-mini …but maybe my 3.7 experience is polluted by using it in 0.46).
2cents on why.
Focused/small iterations - ai tries to focus just on what you want it to do and reasonably build upon existing work. Your “coding style”
0.46 - the ai suddenly started building rocket from a boat.
….that would be ok, if it did not produce so many bugs.
Fixing bugs in others ppl code is a nightmare.
…..so now you are fixing bugs in large and complex AI code.
AI code which is most likely levels above your personal level of coder development in terms of style.
0.45.11 & 0.45.14 and DSR1 nails it for 85% of the time. Make sure to start new chats after about 3 or 4 questions, good rules and Cursor will act as your best companion as you’re used to. Claude sometimes gets it right when DSR1 starts to fail, or C starts to get hickups and forget to show the R part or apply button. Then I just start a new chat, again.
When generating projects, I usually use the 3.7-think version. For making revisions or local adjustments, I use 3.5 since it follows instructions better. Everyone can try this approach.
Good for you anonymous chatbot. My experience is that since 45 it got worse and worse. Tested them all, nearly every day. Started, tried for 60+ min, closed frustrated. Hardly had fun with cursor in 2025. Was about to leave. With this tip ill stay
Anecdote:
Maybe we are all mass hallucinating but I totally agree with this thread’s sentiment. Not an angry customer here. Very happy with my cursor usage up until the downhill tread of 0.46 and henceforth. It’s difficult to point to what exactly is going wrong and I’m sorry that I can’t articulate it to the team better. LTS of 0.45 would be the only path forward with this product imho. Everything else is on the wrong side of history.
Bug Report:
I reverted back to 0.45.17 after taking 0.48.x for a spin and none of the models can be chat with anymore. How/why did this break?
