Did something change? Did the models chosen behind the scenes change?
Auto mode, at least up till a couple of days ago, was not too bad. It would actually work and solve problems and do things.
However, as of now, it is abysmally bad, to the point of effectively being useless. It will not even remember things that were said a couple prompts back and will actively sabotage itself as if it just can’t remember or can’t reason about something it’s actively working on.
For example. Adding a .env file. It doesn’t add the .env file, then tries to load it. Then gets confused, and tries to load it from the repo root. Then gets confused further and tries to create workarounds for env files not loading …etc
I’ve noticed that almost every prompt that I do through auto today has resulted in more problems than it has solved. It’s abysmal.
I made this post yesterday but didn’t actually post it, thinking maybe there’s a bias here, but even today auto is almost entirely incapable of even small, trivial, changes.
cursor updates are like a box of chocolates, you never know when its gonna taste like ■■■■■■ honestly, the last two patches they pushed i think increased the responsiveness and logic for auto mode. We also just had gpt 5.4 launch so 5.3 is more commonly used now and claude also had a big push. we kind of benefit when those two continue duking it out. The last update (today) was not great and it did cause auto to stop following direct instructions and rules… give them 24 hours and they will fix it, for all my gripes Cursor tends to come through and they still have the best tool on the market (ive used them all)
try using markdown files in lieu of auto mode, give it solid guardrails to stay within. You could use Ollama and if you have solid guardrails and a good .md it will perform better just as good as the context you give it
I just updated cursor and I’m experiencing the same, it’s so bad that I came here to check if someone else had this problem, it’s completely useless as it is, I have undone everything it has tried to do. It was working okay before the update
I came here also to look for posts on this. I have been experiencing this for almost 2 days now, and was hopeful that the install earlier today would address the issue, but it almost made it worse! It is dog slow, it does not seem to understand basic asks, and it is scary when I see what it is “thinking” (seems to be really confused). I should note that I have turned off GPT** (for a while now) and have enabled only Claude 4.6 based models.
That seems a bit confusing. I would have thought it would be limited to the selected models in Cursor settings. Anyway, good to know (and still curious as to what is going on - I see a little bit of improvement since last night).
Yea, the description of Auto should emphasize that is has nothing to do with your model selections. If it did, I would just uncheck everything but Opus, if it was going to charge me Auto rates.
Just to test, I explicitly switched to Opus 4.6 (Thinking), and the difference is pretty stark! Again, I mistakenly thought that Auto obeys the models selected in Cursor settings, and I wish I had switched to it in the chat sooner. Edit: Of course, API usage for this is through the roof. I may have to revert to Auto. Sigh!
Opus is ridiculously expensive, but its good. I usually only use sonnet if I am doing something particularly complex, and usually non thinking.
I think we are seeing a pattern across AI coding agents everywhere: the price is going to increase. Whether that is through explicit price increases or gradually reducing function of the cheaper/free solutions, like what Cursor may doing with Auto. Eventually Auto could be pointless to even use for most people, and people will just have to pay for premium models or code themselves. I was using Auto instead of Sonnet, and I had to babysit it so much and fix things and micro manage, makes me wonder if I would just be faster programming it myself. I still think it was sorta helpful, but yea, you can tell the difference for sure.
AI coding agents have always been too good to be true, and we were (still am) in the honeymoon phase of AI agents, where every company was/is racing for early adopters and market saturation, before profits. Now that people have become reliant and can’t go back, the price will increase undoubtedly. Luckily I can program, but as soon as my projects start earning money I won’t mind dropping $500/mo on AI, way too helpful esp for solo developers.
Yes, this is what I was thinking as well. Completely agree it has been too good a ride so far (something that is easy to overlook or forget as we tend get used to any new technology rather quickly).
I have a rule where the model tells me what it is. It is not 100% accurate, but this is the first time I have ever seen Auto say it was gpt 5.4 nano. Which is a very low performance model. I know my request was simple, but just something to note. And I know it was not the usual models I get when using Auto because the way the model behaved was different (more rapid messages instead of one thinking message and then a solution).
Can you share this rule? I am very curious about it. I’ve tried asking the model what it is, and it provides no information indicating what it actually is.
I have also noticed the auto agent has become extremely argumentative and refusing to complete tasks that it routinely performed in the past. When switching to a specific GPT agent, my entire monthly pro token allotment (which historically lasted all month) was exhaust in 24 hours.
I’m now considering switching to another IDE. Does anyone know a comparable alternative to cursor? Google suggested Windsurfer.
MANDATORY: At the beginning of every chat response, you MUSTclearly state themodel name, model ID, model type, and its revision (updated date, only the year). This does not apply to inline edits, only chat responses.
This is the first time I enabled on-demand usage and considered upgrading to a higher tier. And that’s really the reason for this degradation. The value of these tools is insanely high and what we are paying for it does not match that value, and they know it. So we are getting nudged into paying more. It was fun while it lasted. I looked for alternatives, and there really aren’t any if you want a decent IDE experience. So these forums are here to let us vent until we come around to understanding that we have to pay more to continue to get the same output we have become accustomed to.
I may have to revert to Auto. Sigh!
