Let’s be real, at this point they’re just coasting. There’s no appreciable improvement to function, model usage, context management, UI, coding, terminal use, or any other useful feature of Cursor. In fact, many of the ‘updates’ have made these worse to use.
They’re just collecting more funding and using us as numbers to justify it to investors. They don’t give a ■■■■ about use experience or comfort, or they would fix any number of major UI and integration bugs that have been talked about since January. They would work on making the agents actually follow rules. They would fix the API key issues. They would allow people to use their own models and APIs as long as they kept paying for the program itself.
The only thing Cursor does better than the rest is integrating the AI features with the actual IDE, and others are catching up quickly.
When I can use a MASSIVE Llama 4 model for $0.83/million tokens on my other tools, why am I supposed to keep paying Cursor $20/mo to struggle and fight with the same old Claude and gpt agents?
Does it look like something as a working solution?
I worked even with half-day fully stoned coders, but they were able to check the to do lists and implement it…they just regularly failed putting back lines they commented out for testing. LOL.
Check the screenshots.
And I was typically the lead or the hired advisor for project/pipeline optimization, put I made many advising for workflows for open source projects, either. It is the first time I use a dedicated app for the coding part instead of humans.
I make planning, specification, the ‘how it should work’ part.
i expect cursor to follow my rules, to what I ask for. It was able on the first 2 days. Now my workflow is that any new function has to be implemented by CoPilot, then ask Cursor integrate it to its own code. Cursor itself wasted 3 days for me, so I started checking if my prompt is wrong or Cursor got dumber. Cursor got dumber. Copilot does everything i first or second step perfectly. Cursor cannot solve even s simple problem, goes to circular suggestions. And I do not know what model i used as trial does not share that data.
As I told, I was amazed on first 2 days. Now I feel it was a ‘love at first date’ feeling and it was wise to wait out, what happens. Well, if it is a not public ‘free pro trial’ fallback (no full functions, so a lie) to make users pay at first day, it pretty much not worked on me.
That’s not the point, and isn’t related to anything I said.
Sure, AI coding is faster than manual coding. I can do that with Cursor, for $20 a month and a painful user experience, or I can do it in other VS Code extensions, AI IDEs, or just something like Claude Code in the same time without the higher cost or painful experience doing so.
Cursor keeps getting lower in quality as they update it to include more functionality. Your premise that these updates somehow cater to the masses rather than power users is not backed by a single shred of evidence, either your own anecdotal evidence or other people’s. Cursor is pushing out updates so fast, that they need to stop taking feedback.
See for instance, Random keyboard interrupts
or the large number of bug reports
This was initially a great idea. Now it is a royal mess of trying to do everything and not getting the basics right - you do not ship half-baked stuff into updates.
Cursor forced me to start playing with local LLMs myself. So depends what you can run locally, but it is as smart as useful as prompt you pass to it. But you will never fit whole codebase in whatever context you get, so you need to be smart about that too. So I started writing AST parsers, and better the context I generated, more useful the local LLM, even useful to the point that I can fit into 4G of RAM something that can tell me exactly (or whoever I want by that) how to implement new feature on my own code.
Which just means Cursor somehow screwed up the middleware between what we type there and what rubbish we get from the LLM. Because I just can’t believe that Sonnet 4 on CC can be this useful, and the same Sonnet 4 can waste so much of everyone’s time and energy in Cursor lately.
Totally agree. I’ve been using auto mode a lot lately and it works great, assuming I plan well and write good prompts.
Whenever I see people trashing auto mode, the quality of their forum posts suggests what their true problem is
I mostly agree with your overall post, but not necessarily about Cursor needing to be more expensive. I think token-based billing is a good approach so that power users will pay more, casual users can still pay a casual price.
The part I heartily agree with is that Cursor should focus on being a tool for serious developers, rather than trying to also be a vibe coder tool.
’ think token-based billing is a good approach’
Token-based billing was always a business/sales trick to make sure the customer never understand how much he has to pay for a service. I never thought it will come back, but here we are.
I could try some, many gives no right to run. And it still does not resolve to see which model is unusable and which is usable with endless circular not working suggestions. Also it would be great to see a pricing on the usage in trial mode (I know it is against your sales interest, but I am a potential customer, so I represent my interests). It could turn out that the way I use it would cost more than hiring a coder. So I lack transparency here.
I don’t necessarily need an itemized list of every request and how many tokens it used. But it’s pretty easy to tell when you pay less per month with token-based billing than with the subscription models.
It also removes the “When will I hit the limit?” fear from bigger projects. I definitely prefer just having straight up usage billing rather than will they/won’t they issues.
I need that. I want to see in the shop which items costs what. Even in cases like today when I got many non-working solutions from cursor… i would like to see how much those constant fails would cost … anyway, it is a business strategy. stinky one.
Would you please share with me your experiences? In worst case I could build up a local system, but I am no coder, i design systems/models with specifications (what comes in, what goes out literally. So for me it is 'create a module gets input from these places, need to process them by these rules and make an output could be wired to the next module in this or that format (oversimplification of course).
The LLM providers (OpenAI, Anthropic, etc) charge based on tokens for using their API. I’m pretty sure all of them do this.
Understand how many tokens you’re using = understand how much you have to pay.
If Cursor doesn’t do the same billing, then it’s going to be very hard for them to manage costs of API usage against their income from some other billing structure.
They do the same stinky business tricks. Think about tokens, they are like an artificial currency. Any time Cursor could show in your console how much you spent with your request. Any time. Same for their providers, of course. It is an old, stinky business trick (3d renderfarms used it for a while) to confuse the users not to see the clear costs. There is technical challenge in it, it is a business decision where the coders of Cursor assist to the business guys to harm their own ‘comrades’ (and long term goals as people forgot to think in long term goals) with the lack of transparency (which is vital for a good and long term business).
Saying token-based pricing is a salesy/business trick is a clear sign that this may not be the tool for you, maybe even any LLM tool, actually. Would you rather pay their electricity bill?
Great take @Stephan, completely agree, been on the Cursor bandwagon since the early days, and would love to see them continue to focus on premiumization.
You still did not get the issue with token schemes in any business, do you?
It is in almost every cases to decline transparency for the customer.
But hey, let it stay token based.
UX-suggestion: make search menu third at size and put a token to your currency converter there instead with showing remaining plan usage/extra money spent/last request cost.
Maybe I should ask CoPilot to code it and send to Cursor team. At least that will be a working code, lol.
‘tool for you, maybe even any LLM tool, actually. Would you rather pay their electricity bill?’
As I am system/pipeline designer there are cases when I would rather pay electricity bill, because that could be: cheaper/faster/more reliable/more secure. And there are cases when a cloud rent is cheaper, also cases when a direct solution provider is cheaper.
The problem is that at this very moment I cannot decide as Cursor hides relevant usage data/costs from me (which model was used for a task as an example).
And you know what, any time I met this token schemes and I recalculated to real cost it came out that there were many more efficient/flexible and cheaper methods to build, saving for companies 30/40% in costs per project with more flexibility. electricity bills included. Huge ones.
My real problem is that I use agents to make the code and I do not know which one could do coding for me by prompted specifications/fixes.
And I do not know which model or environment could be a solution as I do not know which models were used on my first 2 days when I was amazed by Cursor and which are used now (pretty much lobotomized ones, not keeping rules, etc, I used auto).
Aren’t all programmers also system/pipeline designers? I guess it is the case that not all system/pipeline designers are programmers? If that’s true then what is a system/pipeline designer? Is that a more elegant way to say middle-manager/task master? Would that make you a checker instead of a doer?
