Has auto mode become worse quality?

Ideas Feedback
,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I’ve been a long time high volume cursor user.

In the past 2 months, Cursor’s auto-mode has become about 50x worse.

Context memory is horrible, adherence to rules is non-existant 60% of the time, overall the intelligence of the models on auto is nearly unusable compared to what it was in January.

Did I break something in my cursor rules or directory structure, or has there been scale backs of the models selected in auto-mode, and cost-savings done on cursor’s side?

I’m very disappointed to the tune of looking to terminate my annual subscription and move to other platforms.

I have had amazing support and experiences with Cursor and the Cursor support team, so I am just looking for input and an idea of how things have changed, that may no longer suit how I use cursor. If you tell me that not much has changed, then it is definitely on my side. Just looking for some clarity as I’ve reworked my directory and rules stucture 3 times now and restored old set-ups to no avail.

I do not use it for coding anymore, but rather as a directory based structure for interfacing with knowledge bases and agent style research, development, and creative tasks.

Thank you,

Steps to Reproduce

N/A

Operating System

MacOS

Version Information

n/an/an/an/an/an/an/an/an/an/a

For AI issues: which model did you use?

auto

Does this stop you from using Cursor

Yes - Cursor is unusable

6

Auto isn’t a fixed model, and it isn’t tied to a Cursor version. It’s a router that picks a model per request, balancing intelligence, speed, and cost. The set of models behind it shifts over time, and for a lot of requests it’ll choose a faster, lighter model (Cursor’s Composer or a Flash-class model). That’s a solid default for everyday coding, but it’s a less of a fit for the reasoning- and context-heavy, non-coding work you’re describing (knowledge bases, agentic research, creative tasks). That mismatch is the most likely cause of the “context memory” and rule-adherence drop, not your setup. To be clear on your direct question: the frontier models themselves haven’t been weakened. Auto just leans toward efficient models for many requests by design.

The fix is to stop using Auto for this work and pin a strong reasoning model directly in the model picker:

  • For your use case I’d pin Claude Opus 4.8 or a GPT-5.x model. You’ll get far more consistent reasoning, context retention, and rule-following than Auto’s routing gives you. (Worth noting: you leaned on a pinned frontier model earlier this year, and that’s closer to the experience you’re remembering.)
  • You can see which model handled any given response in the message metadata, so you can confirm what you’re actually getting.

Two things that also help with the “context memory” issue specifically:

  • Keep always-applied rules lean. On long chats, Cursor compacts older context to fit the window, which can dilute large always-on rule sets. Splitting big rules into smaller, scoped ones (and only marking the essentials “always apply”) helps them stick. Rules docs
  • Start a fresh chat for each new task. Very long threads get summarized, which is usually where “it forgot what we were doing” creeps in.

If there’s a specific response where the reasoning went off the rails, grab the Request ID (the “…” menu on the message, then Copy Request ID) and drop it here. I’m happy to look at exactly which model handled it and what happened.

Give pinning a frontier model a shot before making any decisions on your subscription. I think it’ll get you much closer to the experience you had earlier. More on the model options here: Models & pricing

7

For what it’s worth, for mid-level coding tasks I’ve found that Auto has gotten a lot better in recent months!

8

Both very helpful @mohitjain @Daniel_Coimbra , thanks! I appreciate the input and experience from your end.

The work I’m pushing for does not involve much reasoning. The failure most recently is compliance.

I do defensive prompt engineering on my side: rules, directory structure, prompts that say “read these files,” “full pass,” “do not stop early.” Not because the task needs genius reasoning. Most of it is compliance, follow the checklist, do not skip steps, do not call it done until it is done.

That used to work.

The model did not need to be Opus. It just needed to obey the prompt, and auto would do that.

What feels new is Auto optimizing for efficiency and turn completion instead of compliance. Recent examples: I ask for a full pass, I get a template and a stub plus “want me to execute?” I ask for line-by-line truth, I get shortcuts until I push back.

So telling me to pin frontier models misses the point. I should not need frontier models for checklist work. And if I do pin them in Cursor, I pay a premium vs going straight to Anthropic/OpenAI.

Auto was supposed to carry the volume and make that tradeoff worth it. If Auto will not comply on heavy prompts, in the name of efficiency, Cursor loses the main use case for which began using it for.

Daniel’s suggest that mid-level coding tasks have felt much better in recent months, leads me to think that these optimizations are likely optimal for Cursor’s coding users (main intention) and my use-case is being optimized against.

I’m not sure if there’s anything left I can do in the rules or workspace configuration that could optimize further for my use-case, but I will look for user examples on the topics before looking to migrate elsewhere.

9

Daniel’s suggest that mid-level coding tasks have felt much better in recent months, leads me to think that these optimizations are likely optimal for Cursor’s coding users (main intention) and my use-case is being optimized against.

Wow! This is problematic, considering that just this week I have been designing my first “agent harness” for a long-running coding task to build a Python lib following TDD and strict Markdowns. I was indeed thinking that Auto would carry the volume; I’m at 19% Auto and 79% API in my Cursor Pro+ plan, even though I used Auto so much more. I have not yet begun this long-running coding task. Let’s see how that goes…

10

Different models read deeper into directories than others. Sometimes it built in under the guise of progressive disclosure training and other because the models stack is shallow. If subagent rules make them use a model for choosing what to access, the old gpt models, gemini and even the new glm5.2 tend to bail out as soon as they find the first plausible answer, They quit digging. Gemini will often stop at the folder name and infer. Not good for accuracy.

But keep in mind for your use case, cursor has a huge fast indexed cache. What it indexed is determined by, gitignore, cursorignore, and cursorindexingignore. If these are set right, composer tends to use a lot of cache. Make sure the files you are expecting to find aren’t in an ignored directory. make sure you don’t have stale docs going into the cache either. Confusing cheap models doesn’t work well, and the expensive ones, well it just makes them even more expensive. Composer 2.5 for everything, no auto, I consider it obsolete.