After the update, it seems that the results are worse, requiring a lot of back-and-forth communication. I’m wondering if any backend system settings were changed.
Thanks for reporting this! Which feature seems worse - chat, CMD-K, or Composer?
Yes, what happened? Claude 3.5 is much weaker now and required much more tweaking to get the correct result. I am use the cmd + L side chat
what version are you on? If you are on 0.39.6, could you try the newest version and let me know if it is still degraded?
Version: 0.40.1
VSCode Version: 1.91.1 - Yeah it’s terrible, I’m back to cutting and pasting from chat gpt and Claude on the web. I don’t want to waste my runs that I have paid extra for. I bought 2000 this month. What once took Claude one or two times, now it cannot get almost at all and forgets to keep everything and just worse than gpt 3.5 right now.
Do you have an example you can share? would help debug, don’t see anything obvious on our end.
are you using normal chat or long context chat? are you attaching any docs, etc. or just the file?
If easier, feel free to email us at hi@cursor.com!
Normal chat. Not really sure what to send? RIght now, I am doing a fairly easy parsing task in Python that Claude would normally completely understand and get perfect the first or second time now is taking numerous times and then it still does not work.
I’ve noticed something similar since 0.40.1, using composer and Claude 3.5 sonnet; Claude seems noticeably more “lazy” than it used to be and as a result (I think) it’s leading to problems when applying changes. Thought it was just me. That said, I haven’t tried other models yet. If I can get something concrete to show as an example I’ll try.
Yeah, I just tried again, I blew through 10 + runs and it did not get it. Something is definitely wrong. I will test it against Claude on the web and report back.
Yep, Claude fixed it in two prompts on the web and I spent at least the last 20 -30 runs for nothing. (I’m sure you all will be able to tell once you look at my account which was the affected runs) I can send you my account email. Will you all be able to refund or give credits? The script Claude gave me on the web was 100 lines and the all the one’s in Cursor have been 50 to 60 tops, lazy and forgetting or erasing what we already did.
Just went ahead and credited some requests back! It would be really helpful if you could share the chat thread with us (file + messages)!
Could you send me an email to send you and I’ll send you the conversation and the output from the web Claude.
hi@cursor.com or rishabh@anysphere.co would be great! Thank you!
I just sent it. Thanks!
Morning! Any luck on getting things working again?
@rishabhy i’ve also been experiencing a degraded experience all day today
i asked it to changed the width of the progress bar component and it randomly changed the height instead
it has been doing this in composer. it has also been doing this for other components where it randomly changes things i did not ask for and makes breaking changes. I’ve spent all day dealing with this and wasting credits
please advice
Ok, I am back and the degradation of Claude is still there and I have been testing Web Claude’s outputs to Cursor Claude, Web Claude appears to be suffering from laziness as well more than I have ever know it to. It randomly changed plans midstream and started coding something completely different that what I had asked for and and am working on and is only generating a max of 120 lines of code. Cursor Claude does not seem to want to generate anything more that 60 lines and then often does not paste correctly. For the past few weeks before the few last updates in a row, my experience with Cursor has been completely different in that I was able to get to hundreds or a thousand lines of code in a few prompts, and now I am just reexplaining the job over and over again as it makes mistakes and gives the same 60 lines with one change.
I noticed a degradation of Claude as well after the latest Cursor update. I use the Composer primarily.
It seems like Claude doesn’t accurately consider my codebase when making suggestions. It’s quite lazy now. I’ve navigated this by being patient and continually prompting until the code solution is correct.
Chat is what I am using. It appears as it wants to talk more and do less coding.
To be fair there is a degradation on Claude Web as well. Seems like it’s not just OpenAI that’s hit by progressive degradation.
A lot of people are complaining about it. My guess is that Anthropic had to reduce costs.