Finding same thing here. Sonnet 3.7 is horrific today. Tool calls failing. Specific instructions being ignored. Its pretty much unusable for any big task
3 Likes
The previous version of Cursor generated code far more reliably, but now it’s a complete mess. The agent often ignores user intent, and unless you’re using Anthropic’s AI, it fails on too many requests — and the cost of those failures is passed on to the user.
What’s worse is that even with Anthropic’s AI, it modifies too much, wasting even more requests.
The UX has become more confusing, and it forces users into very specific workflows. If this continues even a little longer, Cursor will be quickly overtaken. Personally, I’ve already started using tools like Roo Code and Copilot more than Cursor.
If the competitors release just a few more solid updates, there will be absolutely no reason to keep paying for Cursor. But Cursor keeps making things worse with every update.
2 Likes
This has happened to me, too! Something’s wrong with it. It is taking forever to fix things, if they ever do, and creating problems that I must spend hours refactoring the code. To get it working again.
2 Likes
Happens to me too
@Cursor please fix
1 Like
I’ve used cursor daily for months and something is definitely wrong… I agree with the sentiment. On some simple prompts it is doing the exact opposite of what i request. Hopefully Cursor takes notice, but whatever changed over the past 24 hours needs to be fixed.
1 Like
I’ve noticed some flakiness (ie: deleting a bunch of tests when I wanted it to just update another) but then I ensure it knows to check my codebase by telling it to do that, I also ensure I reference the files in the context.
It helps I know how to actually program so when it does give me bad info or doesn’t do what it should I can still debug and fix. So for me I haven’t really found it to be an issue, and when I do it’s usually something I can just tweak or clarify.
Here’s an example, Dan — even Agent Mode can’t edit the file. What is this format, anyway? Do I really have to copy and paste it manually? This is really disappointing.
Also, Agent Mode doesn’t respect my rules. The rule type is set to “always,” but it keeps ignoring them. Cursor is getting more expensive day by day, yet the experience keeps getting worse.
2 Likes
I am using auto, and results are getting terrible as of today, was working great like yesterday?
Actually its really odd. I have been using this week with great efficiency. But today my god, it was insane. It’s like i was talking to different person at each prompt. Not knowing what was the coding we were dealing with.. Lets hope it doesnt last.
I just started using cursor a week ago and as of last night when I manually select Sonnet 3.7 I get bad code. I check my account usage and gpt-4o-mini or cursor-small uses keep going up instead of premium. If i switch to sonnet 3.5 it starts using sonnet again.
I seems there is a bug where it is not always using the selected model. The difference between 4o-mini and sonnet is pretty huge and it is a little frustrating when you want to use sonnet 3.7 and it is using 4o-mini without any indication other than account credits.
So far I am liking cursor but this is a pretty annoying bug.
2 Likes
I got the similar situation. Gemini 2.5 pro edit code files failed but claude soonet 3.7 seems better
Yes, I have also encountered such problems. Gemini 2.5 pro always fails to edit code files or ends the conversation without editing.