Cursor behavior changed

Ideas Feedback
1

Hello team and Cursor fans,

I’m working with Cursor almost on daily basis since early June now and since a few days I’ve seen a drastic change in Cursor behavior, discussion tone, quality of developments including unattended regressions, previous dev corrupted, local work folder removed, difficulties in understanding instructions, diminished creativity etc…

I was so far totally astonished by the level of collaboration we had. Like a small but precise prompt, instructions for a band new application and this was enough for Cursor to develop in 30mn a full v1.

Now I need to spend/lose time on asking “it” (or him) to repair its mistakes, restore lost data, redo some job lost, or even just revert some commits because of misunderstandings or bad extrapolation.

I truly see that there’s a clear change in the last days, in the quality of the job done, as well as a few language drift in our discussions. Where “he” was super pro, concise, precise and structured, it became messy, few slang words for commenting its mistakes, etc…

This is not a critical report (even if I have material for this) but just a willing to understand what could have happened ?

I’m already Pro+ plan and spending more money on Cursor than Netflix… so I’d love to know when a change is coming or has happened and why.

I’m opened to discussions of course and debate.

Thanks a lot anyway to all, to bringing this amazing tool on the market, a few steps beyond the competition (still stuck to 1 model), and to all other Cursor fans for their comments.

Enjoy

6

Hi @jean-luc Thanks for your forum post! Nothing changed on your account, and this isn’t something you misconfigured. What this likely lines up with is a shift in the model behind your Auto usage. While we don’t disclose exact details about the model used by Auto, in your case, a lot of your requests were moved to a larger, new model. Models differ a lot in personality even when they’re comparably capable, in tone, in structure, and in how literally they follow instructions, so the same prompt can start coming back looser. That’s the most likely explanation for the shift you felt.

If you want consistency, choose a specific model in the model picker instead of Auto. Also, remember that we double included usage for First Party Models (e.g. all Cursor models). So you can experiment with Cursor Grok 4.5 and Composer 2.5 and see which one suits your needs best. These two models draw from the same first party model usage pool that Auto does.

Additionally, I have two small suggestions that may also help. Start a fresh conversation per task, since long sessions get compacted and lose earlier detail. And commit regularly. Checkpoints let you hover an earlier message and click Restore Checkpoint to roll files back within a session, but they’re local and separate from Git, so they aren’t durable history.

Last thing: the thumbs up and thumbs down on each response are the most direct way to influence this. Thumbs down lets you add what went wrong, and those ratings feed into how model quality is evaluated. I’ve passed your feedback along to the team as well.