I know there are rants about the pricing changes of Cursor, I just wanted to shed light on the matter from a different - and hopefully reasonable - perspective. It is about removing unlimited automatic model usage, and all the consequences it is going to come with.
Just to say a couple of positive words, Cursor have improved greatly in the recent months, the guys behind its development are doing an amazing job.
However, after using their Pro Plus subscription fairly heavily in the recent weeks, mostly relying on their auto mode feature, I don’t have the feeling that every paid-for token is justified, not even with the latest and greatest models, but especially not with auto mode.
Within a bit more than 2 weeks I have used around 800 million tokens just in auto mode, with 864 million in total, and although my auto mode is currently free until the 11th of November, so far my usage would be $325. I assume this number will easily reach $600 by the end of my billing period.
To get to the point, most of my Plan mode and Agent requests fail. There is a noticeable improvement when “auto” selects Claude 4.5, which to my great satisfaction it very often does, but still, most of my finished agent requests fail to achieve their goal. As long as my usage is not billed per token I can live with this, but as soon as these failed requests are going to cost me a lot of money they can’t be justified anymore.
Maybe this is different with other languages, maybe with web stuff or when producing CRUD code it is more efficient, but with C++ development it does fail a lot, it makes beginner mistakes, it introduces bugs, breaks already working code.
So what will happen when unlimited auto is phased out? How are Cursor’s customers supposed to feel when their agent just spent a couple millions of tokens on an agent task without a working result? I don’t think that anyone will just easily pull out their debit card and happily pay for token usage at the end of their $200 included usage in Ultra, when their agent requests are spent on unfinished and failed tasks.
I am on Pro Plus at the moment and I was considering Ultra seriously, but after I saw my usage growing over a theoretical $300 in 2 weeks, with way too many failed requests and bugs, I ended up cancelling my subscription.
I would be curious how others feel about the success rate of agent requests, is it better with web development or other areas? I just recently asked Cursor to create me a basic product page and deploy it to the cloud, and I was surprised that it worked much longer autonomously than it usually does with C++ code, and it tested its results too. So I have the feeling that with webdevs may not agree with all my points here. It would be interesting to hear what you guys think.