Hello. I encountered a weird bug / limit cap with the R1 model… When I have a complex, hard prompt with either one big file (a few hundred lines of code) or a medium codebase (a few thousand lines of code), R1 initiates the process, generates all that inner dialogue, and suddenly stops in about 70% of that inner dialogue without a real answer.
This seems capped by time or length of the answer, because when I get a short answer from R1 - it is the full answer, but when it is longer than X amount of time, then it is stopped in about the same limit.
I tried the WiFi router QoS on and off… I tried VPNs with different countries, I tried Cloudflare WARP, and different DNS… I even tried to share my 4G from my smartphone to my laptop and use this connection. Tried to turn off all VS Code/Cursor extensions. Restart PC. Close all other apps. Create a new chat. Create a new compose. Tried to move the project to another folder. Remove and rebuild the index. Have updated Chrome and all other apps and updated Windows 10. I have the latest version of the Cursor, 0.45.4. I never had this issue with other models.
Don’t have this issue with R1 on OpenRouter with the same files and prompts. I can share in DM/PM my prompt, recorded bug on video, and even my codebase if this helps.
If this will not be solved, I want a refund or pause my subscription until it will be resolved (because other models are useless now compared to R1… I rarely use Sonnet 3.5 on some light tasks and never ever benefit from other models, even o1, at least for coding) and possibly add more days to the account because I lost so many days, not just being unable to use Cursor but also trying to deal with that bug.
BTW, I love Cursor, but there are so many bugs… that I am considering switching to something else (I don’t want to give names here because it will be kinda rude, but one competitor gives $100 free credits for R1 and can be used in open-source IDE alternatives to Cursor. Well, I don’t mind paying for Cursor, but when you have so many bugs and literally free alternative, the choice is probably obvious?).
It feels like alpha version 0.4 of the unfinished app, not like a commercial product (all previous bugs with context, issues with code applying, and so on).
But even despite all of that, I still love Cursor more than any alternatives, but you guys really should fix all bugs faster, especially in the era of AI when everything is moving so fast and competitors are breathing down your necks.
Maybe you should collaborate with open-source devs? It would be a killer!
P.S. I think this is related to Cursor doesn't let R1 think , but I described it in more details.