Cursor 0.46 unable to solve a problem, Cursor 0.45 fixed it in one shot with Sonnet 3.7 thinking

I’ve been using cursor for a while with good results. 0.46 has however caused me a lot of issues for the past few days. I originally thought it was sonnet 3.7 but now I think its cursor.

I had an error

The inferred type of 'auth' cannot be named without a reference to '../../../node_modules/better-auth/dist/types/helper'. This is likely not portable. A type annotation is necessary.

In 0.46 I was unable to fix this despite several attempts with agent, chat and edit (I’m not even sure of the difference between these names, it seems edit and agent are exactly the same), cursor kept trying to make up random imports. This was using sonnet 3.7 with and without thinking.

I switched back to Cursor 0.45 and it fixed it in one shot with 3.7 thinking

I have no idea what you have done in 0.46, but I found it difficult to use and it also gives bad results.

I think you seem to think everyone wants to use agent mode, mostly I like the more toned down compose where I could selectively apply edits. Perhaps the vibe coding people like it as they do simple things but for more complex workflows agent mode is simply not good enough and causes more trouble then it is worth.

I hope you actually read this feedback and take it on board, I’ve tried co-pilot several times and did not get good results, it felt like cursor 0.46 did today. Just random things and a confusing and difficult to use UI designed for vibe coders.

17 Likes

I don’t know what the problem is but I’ve been using Cursor for a year now and this latest .46 version is the worst one yet. It barely works.

You can still download .45 here:

I’ve stared to keep an archive of usable versions locally, so I can revert back when needed.

13 Likes

Thank you. I’m going to get a pro subscription soon but I was hesitating after seeing all the bug reports in here. Now I think I’ll get the subscription and install the older. 45 version.
I actually posted for advice in this forum and got some helpful ones but for some reason my post for deleted.

I feel that cursor should promote users to utilize less expensive models such as R1 , heck even give us the option to choose direct API to Deepseek (cost saving) , not everyone cares about privacy and yes they can have our code , my code is not valuable.

My take is the recent upgrade help cursor save $ on token count , but i feel that totally made any version after 0.45 unusable

5 Likes

+1 的确 但是4.5不适配3.7 体验不好

you can do this

go to cursor settings → models → activate deepseek-r1

Yeah what i’m saying is e.g. charge 1/3 credit for R1 usage so people dont depend on Claude so much since its expensive, R1 can be really cheap if you’re using direct API and we can have an option to use the direct API or not , that would save cursor alot of $

2 Likes

Trae 好不好? 中国也在用Cursor吗?

Is the agent working as well in 0.45 as 0.46?

Also, is there a difference in error rate that you know of? Latest 0.46.8 has looots of failed requests, and the agent just stopping out of nowhere and I have to write “continue” or something.

1 Like

The experience with 0.46.8 has been so bad I went back to GitHub Copilot.

Even though Claude Sonnet 3.7 feels faster, eating through your quota, the quality Cursor AI gives you, is some much worse that you end up saving your time when using 3.7 via Copilot.

2 Likes

When Trae can actually remote SSH, it’ll take off.

LOL , what are you talking about
Use
Remote - SSH for Trae

I only dev from SSH , 100% of the time, at first it wouldnt work but it’s perfectly ok now.

You can also ping their devs up , we’re in talks to see if there’s a need to increase context length on trae

LOL ok, yeah let me install something that was just add a few days ago with 200 downloads. Yeah no.

came to say here, is it me or the AI is so dumbed down, the free deep seek chat on browser does miracles in comparison to the one in the app, we are literally downgrading for money are we not?

2 Likes