I still don't understand how cursor is making any money at all. I spend so much time inside cursor, that I am spending 10-20$ per day on additional requests. Now if I connect model provider APIs to windsurf, I'd be spending upwards of 100$ due to amount of tokens I use through the API per day. And if I connect my own API key to Cursor, I immediately get rate limited for any request, because I go well above 50 per minute. And I did try claude code, but its just not on par with my experience with Cursor.
I could probably go much lower, and find a model that is dirt cheap but takes a while; but right now the cutting edge (for my own work) is Claude 4 (non-max / non-thinking). To me it feels like Cursor must be hemorrhaging money. The thing that works for me is that I am able to justify those costs working on my own services, that has some customers, and each added feature gives me almost immediate return on investment. But to me it feels like the current rates that cursor charges are not rooted in reality.
Quickly checking Cursor for the past 4 day period:
Requests: 1049
Lines of Agent Edits: 301k
Tabs accepted: 84
Personally, I have very little complaints or issues with cursor. Only a growing wish list of more features and functionality. Like how cool would it be if asynchronous requests would work? Rather than just waiting for a single request to complete on 10 files, why can't it work on those 10 files in parralel at the same time? Because now so much time is spend waiting for the request to complete (while I work on another part of the app in a different workspace with Cursor).
> I still don't understand how cursor is making any money at all.
They don't make any money. They are burning VC money. Anthropic and OpenAI are probably also not making moeny, but Cursor is making "more no money" than others.
For OpenAI: short answer is no. From what I've seen, their biggest expense is training future models. If they stop that (putting aside the obvious downsides) they'd still be in the hole for a few billion dollars a year.
edit: Well, if they shed the other expenses that only really make sense when training future models (research, more data, fewer employees ..) they would be pretty close to break even.
The market for AI-assisted development is exploding and token costs are plummeting all the time. It makes sense for them to subsidise usage to capture market share in the short-term with the expectation that servicing their users will cost them less in the future.
There is no loyalty or lock in though. There is little real uniqueness. And everyone in AI is trying to make everyone else on AI the "commodity complement"
It's like a horse race.
But yeah enjoy the subsidies. It's like the cheap Ubers of yesteryear.
This is exactly it. Selling the output of a LLM is going to an incredibly cut-throat and low-margin business.
The more interesting, novel, and useful work you wrap the LLM in the more defensible your pricing will be.
That said I think this can describe a lot of agentic code tools - the entire point is that you're not just talking to the raw LLM itself, you're being intermediated by a bunch of useful things that are non-trivial.
I see this with Anthropic most - they seem to have multiple arms in multiple lines of business that go up the abstraction ladder - Claude Code is just one of them. They seem to also be in the customer service automation business as well.
[edit] I think a general trend we're going to see is that "pure" LLM providers are going to try to go up the abstraction ladder as just generating tokens proves unprofitable, colliding immediately with their own customers. There's going to be a LOT of Sherlocking, and the LLM providers are going to have a home field advantage (paying less for inference, existing capability to fine-tune and retrain, and looooooots of VC funding).
This may be old fashioned thinking and the automated loom might come get me but I think traditional software products with enthusiastic customers, some kind of ecosystem will benefit with AI being used.
However they will benefit in a way like they benefit from faster server processors: they still have competition and need to fight to stay relevant.
The customers take a lot of the value (which is good).
While there is a lot of fear around AI and it's founded I do love how no one can really dominate it. And it has Google (new new IBM) on it's toes.
It's hard to add sophisticated abstractions though, because they are all selling text by the pounds (kilos?). So it feels the same as vendor lock for a cucumber seller, doesn't it? The seller can sell you an experience that would lock you in, but aside from it there is no moat since anyone can sell cucumbers.
To try and give examples: an autonomous agent that can integrate with github, read issues, then make pull requests against those issues is a step (or maybe two) above an LLM API (cucumber seller).
It doesn't take much more of a stretch to imagine teams of agents, coordinated by a "programme manager" agent, with "QA agents" working to defined quality metrics, "architect" agents that take initial requirements and break them down into system designs and github issues, and of course the super important "product owner" agent who talks to actual humans and writes initial requirements. Such a "software team system" would be another abstraction level above individual agents like Codex.
When people talk about how sophisticated hierarchical agent swarms will be built up that perfectly reflect existing human social structures I'm reminded distinctly of all the attempts to build DDD frameworks for modeling software, and then the actual result is that software went in the opposite direction - towards flattening.
As native LLM task completion horizons increase another order of magnitude, so much of this falls out of the window.
This exactly. I built CheepCode to do the first part already, so it can accept tasks through Linear etc and submit PRs in GitHub. It already tests its work headlessly (including with Playwright if it’s web code), and I am almost done with the QA agent :-)
My second bet is on Google (for general-purpose LLMs in general) - not because of any technical advantage, but because they have a captive audience of large organizations using GSuite that would be happy to just get Gemini on top to satisfy need for AI tools, instead of having to jump through the hoops of finding another provider. Sales, sales, sales.
Do you mean AWS? They're competing with half a dozen or more hyperscalers now. Cloud infrastructure components are so heavily commoditized now, many of them have open source solutions with compatible API's. (Think Minio)
There is a loyalty if they keep winning, if they stop running their competitors will beat them. I don't switch between cursor or windsurf daily, i keep cursor only even if windsurf has some marginal improvement in workflow as i know cursor will have them in short time. No need to switch, But if they stop improving they will get eaten away. They have already taken lot of developer market share away from vscode and vscode copilot.
Inference cost is plummeting. It’s like the cheap Ubers of yesteryear, if the cost of hiring a driver dropped by a factor of a thousand in the past three years.
I use Aider with Openrouter and I keep wondering about the pricing of LLMs after providers decide to be profitable. Can we still afford a model which knows Python, Java and how to disrupt snail biology without poisoning mammals?
Yes. It’s already profitable to run inference at today’s prices. AWS isn’t subsidising you when you buy compute from them. And inference cost is declining steeply.
> The cost of LLM inference has dropped by a factor of 1,000 in 3 years.
But, we need a future where unlimited inference, in parallel is profitable. It is not: even less than cloud compute (where it is terrible also), when I buy 500 flimflams for $50/mo, what did I buy exactly? As currently it seems to depend on the position of the moon: one time 10 prompts make what I want, sometimes 100 prompts keep looping over the same issue unable to fix it (like a typescript type issue which takes me 1 seconds, llms, the flagship ones, can easily burn 100 prompts and not fix it). I do very much NOT want to pay for those 100. I see 'vibecoders' aka people who cannot code, burn through all Tokens for the month without having anything working in a single day.
The answer to that depends on when the VC bubble bursts- if it lasts long enough costs will eventually drop far enough. Pets.com was a .com-boom era joke but today I actually buy my pet-food online and I'm pretty sure nobody is subsidising me doing that.
Commercial animal feed is subsidized. So are some forms of human food in many countries.
Pet food is not subsidized in my country nor the EU. If any countries do subsidize pet food, they are the exception. Maybe the US? Pet food is often manufactured from the waste of other processes, including the human food industry, but that is not a subsidiary.
That depends, there are specific diets for, say, cats, and it not only needs to be prescribed (to be able to purchase), but it costs a fuckton of money.
<rant>
Think of it like this: imagine if lactose-free or gluten-free food could be bought only with a prescription. Sadly the prices are already high as it is for gluten-free, but I would rather not get into the reasons here. :)
My girlfriend (LA, US) just left 1k USD on 2 visits to the vet with her cat, for some cheap ass antibiotics, and a "specific prescription-only food". Crazy. All that would have been "free" (not quite, but you know) or at a very low cost for humans around here, in Europe. Not the gluten-free food though!
I understand that it is not directly subsidized. However the sources it comes from while are the "waste" of a greater product. That greater product is heavily subsidized.
This also goes to a personal issue that why would you feed your pet a waste product. My dog gets food I cook for him just like myself. There are tons of crock pot recipes online for safe cheap high quality dog food.
> It makes sense for them to subsidise usage to capture market share in the short-term with the expectation that servicing their users will cost them less in the future.
Switching costs are zero and software folks are keen to try new things.
The time it would take me to switch IDE and work process and learn the best prompting style and idiosyncrasies of a new model (and do some testing to build confidence) would be half a day, at very least.
That makes the opportunity cost of switching significant.
This won't create race conditions all of them will know the others live accept of commit to directory? or have to wait then hit enter on other tab instantly after?
they say they are making hundreds of millions, but they never say how much of it is going to GPU cost. If I had to guess, they are burning everything and far from being profitable
If history has taught us anything, it's that unless full accounting data is released, there is a reason that full accounting data is not being released, and that reason would almost certainly paint the company in a bad light.
GPU type and utilization mean that the costs likely rise only logarithmically or sub-linear. If you commit to buying enough inference over long enough, someone can buy a rack of the newest custom inference chips and run them at 100% for you, which may be a lot cheaper per request than doing them on a cpu somewhere.
I meant you're using Cursor (I assume Claud as the model) and also wondering that you go over monthly usages and then what product/project is that where you find it so useful that you have such a usage.
I have used or rather use Claud with CoPilot and I find it pretty useful but at times it gets stuck in niche areas.
I love Cursor, but it feels like a ticking time bomb with extensions not being updated at the same rate as VSCode.
Also another issue I am starting to see is the lack of shared MCP servers. If I have VSCode, Cursor, and Claude open, each one is running its own instance of the MCP server. You can imagine that with a dozen or so MCP's, the memory footprint becomes quite large for no reason.
I think about this daily. More devs are starting to pick up on Claude Code. The initial “not an IDE!” scare is usually diminished within the initial session.
I don’t think the future of agentic software development is in an IDE. Claude Code gives me power to orchestrate - the UX has nothing to do with terminal; it just turns out an agent that lives on the OS and in the filesystem is a powerful thing.
Anthropic can and will evolve Claude Code at a pace cursor cannot evolve IDE abstractions. And then yea - they are designing the perfect wrapper because they are also designing the model.
Note: the text of that article itself is AI generated.
> Automate Content: Like this very post. I use Wispr Flow to talk with Claude, explain the topic and tell it to read my past blog posts to write in my style.
Now I have the mental image of the owner of that blog tearing his hair out trying to get back into his computer, while the AI that locked him out is happily posting on his blog trying to convince other gullible humans to hand over control of their computers.
How are people using Claude Code day to day without spending a lot? I tried it on a moderately complex task and it chewed through tokens at an alarming rate. I quickly spent $2 and hadn’t even arrived at an adequate solution yet. I’ve heard other people say they’ve spent $10-20 in a coding session. I don’t see how that’s sustainable for me, so I’ve stuck with my $20/month Cursor subscription.
Pro isn't a static plan. Pro subs can access Claude Code but are paying via API metering. I have it setup at home and, while I haven't used it much, it can quickly add up.
What I did do, because my personal projects aren't too complex, is moved the model from Opus to Sonnet which is about 1/5 the cost.
For day-to-day stuff I have ProxyAI (on IntelliJ, continue.dev works for this too) pointed at Mistral's Codestra for auto-complete and to Claude 4 for chat.
Claude Code is just for giving the bot tasks to do without me being deeply involved in the work.
Normally to make a new smallish feature it costs me about $0.40.
The core suggestion is to point specifically at the files you want it to read and use as a reference, otherwise it might go read some huge file for no reason. Also the tokens used depend on the size of the project.
Generally, if I'm doing something I can box, I'd use chatgpt and copy it in myself. If I want something matching the style, I'd use a guided Roo Code.
Until Claude Code becomes manageable price wise, I don’t think Cursor really sees them as their competition. I can burn the whole cursor subscription price in a single day with Claude Code.
Ha, same. How do you use it? I tried all the fancy context management stuff multiple times, but I mostly just have a chat buffer open and copy paste stuff manually. Text wrangling is so damn efficient in Emacs. I pay around 10$ to Anthropic per month in API tokens for pretty heavy usage. With deliberate context management (I found keeping it small and focused vastly improves responses), cost is really not an issue.
Didn't try anything agentic within Emacs yet, don't find that helpful enough so far.
> tried all the fancy context management stuff multiple times, but I mostly just have a chat buffer open and copy paste stuff manually.
As of last week you can insert a link to a plain-text file in a chat buffer to include its contents in the prompt. It must be on a line by itself. In Markdown it looks
[like this](/path/to/file)
with Org links in Org chat buffers.
This feature is disabled by default to minimize confusion. To enable it you can flip the header line button that says "ignoring media" to "sending media". This works for sending images and other media too, if the model supports it.
I'm surprised nobody is mentioning how cheap copilot pro is. $20 and you get all you can eat inference without using your own api key for the models on vs code agent mode.
Copilot pales in comparison to Cursor Pro. I've trialed it three or four times in the last two years and stopped using it after a few days each time. Honestly, I have no idea why anyone pays for it given the alternatives.
My only wish is that Cursor had partnered with Zed. vscode isn't enjoyable.
I’m on Pro+ and get rate limited heavily. 1-2 hours of semi heavy use and the brakes kick in. I can’t stay productive in it because this always rips me out
Yeah, if you're a heavy user of Claude code, you pretty much need to use it with a Max subscription rather than a BYOK approach. But that starts at $100 / month so it's a significant bump from cursor.
For any professional SWE 1/200 a month is basically nothing in terms of the value it delivers. They just rolled it out to the 20/month plan with limited usage but as soon as people get used to it I have no idea why they wouldn't upgrade unless they are just entering the field and don't have a job yet.
> For any professional SWE 1/200 a month is basically nothing in terms of the value it delivers.
If it is delivering that value indeed, then 100-200 dollars each month is exactly what that professional SWE is worth.
That SWE is going to have to pivot into something adjacent to provide value-add, and if the only value-add is "I can review and approve/reject", then it's only a matter of time[1] that there are no SWEs who can review and then approve/reject.
And that is assuming that the LLM coding performance remains where it is now, and does not improve to the review, orchestration, design and planning levels.
Anyone pulling that a month not working for themselves doesn't have to think about the cost. That kind of salary is paid by corps with strict privacy policies.
Unless you do nothing else with your time I'm not sure how you'd utilize the $100/mo plan fully.
the pricing is for token usage in 5/hr windows not monthly caps. if you use it intensely a couple times a month within a 5hr window it's not hard to hit the cap and want to upgrade. Personally I just work on some side projects during work on another monitor and just every half an hour so throw something at it and that's been very valuable for me.
225 messages every 5 hours? You hit that on the side while you're doing your day job? I suppose if you push all work to Claude and do nothing else all day it could be a concern but I don't think it would be a very effective way to work in it's current state unless you want to be left with a giant mess.
I admit their transparency around limits is a bit questionable. Why don't they just list out the tokens and remaining time?
Yeah I hit it again today, refactoring can use a ton. It did make a bit of a mess but planning it out making sure everything is tested and having it run through checking the tests making sure they all still pass uses lots of tokens but passively doing that while I'm working is way faster than doing it manually.
Sometimes I'll just experiment with weird stuff and end up reverting it or delete it afterword's. Also fun to build really nice data visualizations for data I'm working with.
> Pro ($20/month): Average users can send approximately 45 messages with Claude every 5 hours, OR send approximately 10-40 prompts with Claude Code every 5 hours. [0]
One of the first features we added to FastMCP 2.0 [1] was server composition and proxying (plus some niceties around trying to reuse servers whenever possible) for exactly this reason, so that you can ideally run a single instance of your preferred servers without replicating the setup over every application. In some ways, MCP standardized how agents talk to APIs but introduced a new frontier of lawless distribution! This is something that has to improve.
Have you tried the Jetbrains agent, I think it’s called Jennie?
I am trying it right now and it seems decent enough but I haven’t tried Cursor as I don’t really like vs code.
Jetbrains Junie is pretty good and comparable to Cursor in my experience. And since it is already included in my Jetbrains license, I have had no need for Cursor.
I don't understand. Is this meant to run locally? Because I tried to deploy my agent using GitHub MCP server to K8s. I can't ask my agent to run docker command in a pod.
Cursor and other VSCode forks connect to Open VSX [1] for extensions. Barring some of the Microsoft extensions, I've found that pretty much all the extensions I use are available and kept up to date on Open VSX. Cursor seems to have enough funding to support their own variants of the Microsoft extensions, like Python and C++.
The one issue I've run into is that the VSCode version Cursor uses is several months old, so we're stuck using older extensions until they update.
for me, I’ll keep doing work on the codebase in a separate vscode while Cursor’s agent is wiling away at it, so as to not be distracted or interrupted by the agent activity in the corner of
my screen. and then i’ll have a claude or aistudio tab open doing bigger analysis or planning tasks, reading papers together, etc.
What would you say to somebody who doesn't trust Cursor to "take the reigns off" and go "agent mode" and do such large changes? Even with "checking/approving/having the final say" as a user, I feel there is time lost if the AI does not do the changes the right way/makes small pesky mistakes.
Autocomplete: yes
Asking questions: yes
I know everybody is all into agent mode and it isn't a Cursor specific thing, I'm just confused why everybody is into it.
I just accept that when I return to the Cursor window, it is going to have made some mistakes and I’m going to have to spend some time fixing things. I agree it’s a delicate balance, and sometimes it feels more like I’m exploring methods of steering the AI than I am doing anything about the code.
my usecases have been building relatively self-contained, well-specified components of distributed systems out of academic papers and my own design notes / prototype skeletons, in rust. there’s a lot of context for the agent to ground against, and it can blow out ideas into their implications relatively well ime.
the experience of fixing up lots of pesky mistakes that you yourself would never make is kinda annoying.
Lots of comments here are so unrelated to the release, just people complaining about how they don’t like cursor because it’s a fork of VSCode…
I’m particularly interested in the release of BugBot. The docs mention it looks at diffs but I hope it’s also scanning through the repository and utilizing full context. Requesting copilot to do a review does the same thing but because it’s only looking at diffs the feedback it provides is pretty useless, mainly just things that a linter could catch.
I keep switching away from and back to Cursor (mainly due to frontier models explicitly fighting their apply model, the first few times it’s funny to see the LLM itself write “this is frustrating” but I digress).
And every time I find it having diverged further from VSCode compatibility.
This wouldn’t be so bad if it was an intentional design choice but it seems more that Microsoft is starting to push them out? Like MS Dev Containers plugin is still recommended by some leftover internal, but if you install it you get pushed to a flow that auto uninstalls it and installs Remote Containers by Anysphere (which works differently and lacks support for some features). And I end up rebuilding my Dev Container once more… I also noticed recent extensions such as the Postgres one from MS also doesn’t exist.
I fully expect MS to change the VS Code license in the not so far future to make applications like Cursor not possible. Forking might be a thing initially but will quickly fade since without the backing of MS the ecosystem around it will die.
This is why I’m using Zed now, and Claude Code. I like to keep Zed pretty minimal and I’m slowly weening off of Cursor in favor of Claude Code when I need it
Yeah, it took me a few weeks to wean off cursor but I’m now happily using Zed exclusively.
Cursors tab predictions are still a bit better and snappier but I feel like Zed is a better editor experience over all and I don’t rely on AI anyway. Agent mode works pretty well for me though. Also cursor leaks memory pretty bad for me.
There’s still room for improvement but Zed is working on fixes and improvements at a high pace and I’m already pretty happy with where it’s at.
I’m not a fan of the 20 tool limit unless you use the Max option which costs you 1 credit for each and every tool call + message. Seems like an artificial limit and it always rips me out
Not likely. They open sourcing the Copilot UI is the way to kill further attempts to fork. Now you don't have to fork to have features you could get only by forking and maintaining the fork. The amount of work to make a Cursor competitor is significantly reduced.
If you pay attention to VSCode changelog for the past few months, you'll notice that most of it is about Copilot.
It feels almost as if VSCode is not adding new features and is in maintenance mode for now. I have no idea if that's actually true, but if this continues, a fork will be easily maintainable.
I must mention https://ampcode.com/manual which is my favorite toy of them all right now. Has almost no settings; uses Claude 4, no way to change model; just works! (Unfortunately, not a paid comment.)
Yes, something like that. As probably many of you, I spent way too much time switching between agentic apps, switching between models inside these apps, and tweaking system prompts for these models.
When Ampcode took it all away from me, I found I enjoyed the actual AI-assisted coding much more than configuring. Of course, largely because it just worked. Granted, I had enough experience with other AI tools to manage my expectations.
The vibes are on point... I guess looking like Perplexity, right?
But for the way I'm using Cursor, I guess the flow is harder? I want to make smallish edits (review code, improve function)
A small thing, but I appreciate their free trial usage "starter pack" - the landscape of SWE assistance tools is pretty large these days, and it's impossible to assess fit for the use cases you're interested in without trials. As much as all the positive feedback and/or hype about Anthropic's product is tempting, for hobbyist use I can't quite justify shelling $20 out of the box. (That's to say I can't bill it to someone paying me...)
When reviewing the changes made from agent mode, I don’t know why the model made the change or whether the model even made the change versus a tool call making the change. It’s a pain to go fish out the reason from a long response the model gives in the chat.
Example: I recently asked a model to set up shadcn for a project, but while trying to debug why things looked pretty broken, I had sift through a bunch of changes that looked like nasty hallucinations and separate those from actual command line changes that came from shadcn's CLI the model called. Ended up having to just do things the old fashioned way to set things up, reading the fine manual and using my brain (I almost forgot I had one)
It would be nice if above every line of code, there’s a clear indication of whether it came from a model and why the model made the change. Like a code comment, but without littering the code with actual comments
That's still not what they're asking for. If it's a commit message, it's already saved to disk. This is at the point where the LLM has written code but the diff to the file hasn't been saved yet.
Even if you were willing to deal with your agent making commits and having to fiddle with git to break those up or undo them or modify them, it still doesn't solve OP's need. They want to know what bits of diff apply to which thinking tokens. The prompt and the output are necessary, but there's no mapping between the final diff and the stream of tokens that came out of the LLM (especially in a commit). But that information is known by the tooling already! When generating the diff, the LLM had to output code changes. So you can map those changes to where they originated from. It's not just "what did I tell the LLM to get this change" it's "what was the thought process that the LLM went through after getting my prompt to end up with this line change"
this is remarkably lowkey for a 1.0 of a big product.
where is the splashy overproduced video? where is the promises of agi? where is the "we are just getting started" / "we cant wait to see what you'll build"? how do i know what to think if you aren't going to tell me what to think?
VSCode with extensions Copilot [autocomplete] + CLINE [AI chat] + FOAM [obsidian-esk markdown support] is goat. There's no way a closed-source alternative to going to compete with this.
The benefit of Aider is that you can configure a very involved static analysis toolchain to edits which directly triggers new edits in response, and everything is a git commit so it's easy to revert bad edits quickly. I have used both and I find Aider provides more control and produces code faster due to leaner prompts (it's also easier to run multiple Aider instances than Cursor instances), while Cursor has a prettier interface, and I do like being able to see diffs live in files (though I almost never spend the time reading them to accept/reject). I imagine if you don't spend any time configuring Aider cursor would probably seem far better.
Create a file like conventions.md in the root of your repository with specific commands for common tasks: running tests, linters, formatters, adding packages
Set this as part of the files it reads on startup. Then ask aider to look at your codebase and add to it :)
Aider has a lot of slash commands to familiarize yourself with. Ask and web are crucial commands to get the most out of it.
My recommendation to anyone is to use ask the most then tell it to “implement what we discussed” when it looks good.
The biggest thing is to set it to autofix lint/test issues, then to set up a really good lint/test config. Also, I find that Aider's default system prompt setup is a little less preconfigured out of the box than Cursor's, so it helps to have detailed styleguide/ai rules documents that are automatically added to the chat. I usually configure my projects to add README.md, STYLEGUIDE.md (how to structure/format code) and AIRULES.md (workflow stuff, for instance being socratic with the user when requirements aren't clear or the prompt contains ambiguity, general software engineering principles/priorities, etc).
I was not commenting on the vintage of this, I was commenting on my ignorance which is where the parent or grandparent comment started from. Sorry if it did not come out right
Copilot in VSCode has autocomplete and also something they call "next edit".
In my experience, next edit is a significant net positive.
It fixes my typos and predicts next things I want to do in other lines of the same file.
For example, if I fix a variable scope from a loop, it automatically scans for similar mistakes nearby and suggests. Editing multiple array values is also intuitive. It will also learn and suggest formatting prefences and other things such as API changes.
Sure, sometimes it suggests things I don't want but on average it is productive to me.
Cursor does this. And in my experience it gets it perfectly right 95% of the time or better. A lot of times I can start editing something and then just keep hitting tab over and over again until the change is complete--including jumping around the file to make edits in various disconnected places. Of course you can do most of this in Copilot too, but you'd expect something that maybe works and needs a lot of cleanup. The cursor autocomplete is, more often than not, EXACTLY what you would have hand crafted, without any deficiencies.
It's also somehow tracking what I look at, because I can look up an API and then return to the original source file and the first autocomplete suggestion is exactly what I was looking up, even though there would be no context to suggest it.
It's really quite magical, and a whole different level from Copilot.
Reading up on this, it sounds like Copilot adopted the methodology that Cursor has been using internally for more than a year. Which is great, but if your question is "why is everyone using Cursor?" it is because many initially used Cursor when they were the only ones with this feature. I, for example, specifically switched from Copilot on VSCode to Cursor because of the spooky accuracy of Cursor's tab complete compared to Copilot, at the time. This was only a few months ago.
Cursor tab is remarkable. There's a lot of competition for agents but I don't think any other product comes close to their tab completion. Admittedly it might be rather useless in the near future with how things are going though.
Interesting. For me Cline and Roo are king. I would use them exclusively if I could afford it. With Copilot Pro+ it goes a long way but still ends in rate limits down the road
I tried Cursor, and will occasionally switch into it, but I'm having a hard time using it because its relationship to extensions (particularly extensions that the user develops and sideloads) is badly broken. I tried doing minor customization (forking the vim plugin from the github version, creating a vscode hello-world-style plugin), and while everything worked in VsCode, transferring those plugins into Cursor did not. There was no documentation for plugins in Cursor, you just had to hope that things were similar-enough to VsCode. And then they failed to load with no debugging leads.
I think this is an artifact of Cursor being a closed-source fork of an open-source project, with a plugin architecture that's heavily reliant on the IDE at least being source-available. And, frankly, taking an open-source project like VsCode and commercializing it without even making it source-available is a dishonorable thing to do, and I'm rooting against them.
I myself run five IDE's in parallel with ~ 6 terminal sessions all on Claude Code. There's a time and place for each tool. Also, thanks to my Neuralink, I have Linear and Azure MCP's wired directly into my frontal cortex at all times. This makes it really easy to keep our docs up to date without lifting a finger.
This product is going to be obsolete with the rise of terminal-based agents like Claude Code and Codex. Why would I abandon my entire workflow and toolset to use a fork of VSCode? Especially one that Microsoft will likely make increasingly difficult to maintain.
Developers use Vim, JetBrains, Emacs, VSCode, and many other tools—what makes you think they’ll switch to your fork?
I wish they had proper support for multi root repos (even though the last update promised better support, it was just a line in the release notes with no docs - which seems to be their usual change management style).
Its so painful - the model never knows the directory in which it is supposed to be and goes on a wild goose chase of searching in the wrong repo. I have to keep guiding it to the right repo. Anyone here has had success with such a setup?
Have you upgraded to the new .mdc file format? I didn't get around to .cursorrules before this format came out, but I'm finding .mdc is reliable if configured well (e.g. with the right file extensions)
My understanding of the docs is that these are all handled the same: Cursor just adds any rules file to the context for each request, and that's it. I don't believe there is any mechanism by which to call special attention to particular rules in the context window. I could try renaming the file though.
Yeah, just putting the structure in the rules and telling it to always specify the full path in commands was enough to fix any multi repo issues for me.
Dev here, Can you give me more details about whats going on here? Screenshots or request ids (or both) would be best
You can email me directly at feldstein at anysphere.co
I've given up on cursor a while back. Too many updates, sometimes if produces shit when you don't need it to react at all, and sometimes for simple really useful autocompletions it hangs for a long time making snippets way more approachable.
Overall, I am having hard time with code autocompletion in IDE. I am using Claude desktop to search for information and bounce off ideas, but having it directly in IDE – I find it too disturbing.
Also there is this whole ordeal with VSCode Marketplace no longer available in Cursor.
I'm not saying AI in IDE is bad, it's just I personally can't get into it to actually feel more productive.
IMO it's a strategic misstep to try and create their own IDE with a fork of VS Code. I'm only going to consider AI Tools that integrate with my IDEs (primarily VS Code + Rider) as such my AI weapons of choice are now: augmentcode.com (fave), GitHub Copilot, Gemini Code Assist and now Claude Code now that I can use it with my pro plan.
AugmentCode is really good. It has mostly replaced my coding for the past 2 weeks. I am "reduced" to prompting, reviewing, and re-prompting.
And I can do this in parallel, working on 2-3 tasks at the same time (using GoLand, AndroidStudio and JetBrains). As long as I can context switch and keep the context in my head.
Yep it's off to a great start with its early mover advantage but IMO their days on top are numbered with every major player behind the premier coding models (and major IDE vendors) iterating hard on their own integrated AI coding agents, after which I suspect Cursor's choice for using a proprietary IDE is going to look dated.
That tells you nothing about their operating expenses (I'd bet they're operating in the red), and if you divide that by their cheapest available plan, that's at most 1.5 million paying users (probably way less).
I stopped using Cursor once Zed released their agent mode. Cursor consistently leaks memory for me even with all extensions removed, Zed is much lighter weight and in my personal opinion just a better editor.
Also, Trae being $10 for more requests makes Cursor far less appealing to me.
The other day, I was repairing a raidz2 cluster. Power instability during a heavy / long term write operation caused problems resembling defective drives and corrupted metadata all over the drives. I came up with a strategy to repair it but I needed some code which was pretty straightforward.
Rust is easier for me than shell scripting so I started writing what I needed and remembered Zed added agent mode. I decided to give it a shot. I had it use Claude 4 with my api tokens.
It wrote the entire program, tested it, debugged it. It made some bad assumptions and I just steered it towards what I needed. By the end of about an hour, I had my complete fix plus an entire ZFS management layer in Rust.
It did cost $11, but that is a drop in the bucket for time saved. I was impressed.
Just sharing this because I got real and measured value recently that is way beyond the widely shared experience.
Yeah, I’m impressed with it too. Sure there are things that can still be improved, especially the tab completion, but I’m very happy with how well agent mode works over all!
As an aside: $11 for an hour of Claude 4 seems expensive? I’ve been using Claude 4 (through a Zed Pro subscription, not with my own keys) and I haven’t hit the subscription limits yet. Were you using burn mode, or?
I don't think it's being kept up to date. I believe for the IDEs, it requires manual testing to get the numbers. Since things change so quickly, it's mostly just a historical artifact. Hopefully some future version is automated.
I've been using Cursor since they merged with Supermaven, but I'm concerned with how they handle controversial feedback on their subreddit.
Recently, there was a post with detailed evidence suggesting Cursor was intentionally throttling requests [1], including reverse engineering and reproducible behaviors. The team initially responded with a "happy to follow up", but later removed their replies that got downvoted, and banned the OP from posting further updates.
Their response sounded AI-generated too, which wasn't very surprising based on the way they handle customer support [2]. I wish they were more open to criticism instead of only claiming to be transparent.
I tried it, it did some magical stuff, like completely rewriting my go test suite from matrix based tests to individual tests. Took a few minutes. Code was mostly correct.
Also had a few misses. But in general it is ok. Still prefer ai assistant, because i can then direct the result into a certain direction. It also feels faster, it probably is not because of the manual stuff involved.
I use it pretty much daily and am pretty happy with it, especially its ability to do very targeted edits rather than leaving random changes everywhere.
I believe it’ll get much better as LLMs start editing code by invoking refactoring tools (rename, change signature, etc.) rather than rewriting the code line by line, as this will let them perform large-scale changes reliably in a way that's similar to how software engineers do them now using IDE tools
my biggest complaint with it is the speed. One of the slowest agents (if not the slowest) I've tried.
It's also missing quite a few features still (like checkpoints, stopping the agent mid-task to redirect, etc). But the core feature set is mostly there.
It behaves weirdly. I opened a Python file in PyCharm and asked Junie: "Tell me about this file". It then proceeds to do about 15 LLM calls, then ends up with a "git diff" and then "Done" (the implementation is error-free). What's the value add here?
Cursor recently lost me as a customer. Too many updates that disturb my workflow and productivity, no easy way to roll back versions, super sparse changelogs, lots of magic in context building, really untransparent pricing on max mode. I recently made the switch to Claude Code on the Max plan and I couldn't be happier. The only real thing I'm missing is the diff view across files, but I assume it's just a matter of time until that's properly implemented in Zed or VSCode.
I feel unstoppable with Claude Code Max. I never thought I'd pay $200 per month for any developer tool, yet here we are, and I also couldn't be happier with it.
I just started with it, so still getting my feet wet, but it's been better than any other tool at really grokking my codebase and understanding my intent. The workflow feels better than a strict IDE integration, but it does get pricey really quickly, and you pretty much need at least the $100 Max subscription.
Luckily, it should be coming with the regular $20 Pro subscription in the near future, so it should be easier to demo and get a feel for it without having to jump in all the way.
Maybe in the US? I will never pay 100$ for a subscription and I despise that people normalized it by even buying this stuff instead of saying "no, that's way too expensive".
I work in a cleanroom to fabricate semiconductor devices and I spend hundreds of euros per hour to use specific tools which mostly just use electricity and maintenance. Should we complain that it’s too expensive or should we use them because they’re worth the price?
Things have a price for a reason. It’s up to you whether it’s worth paying that or not.
We are talking about personal use and then people don't pay for it out of their own pocket but the company's. At least I hope so because otherwise it would be very dumb.
I’m also talking about personal use. These are research devices for my PhD. I’m obviously not paying out of pocket, but my funding agency does.
I’m trying to convey that if a tool increases your efficiency by more than it costs then it’s worth paying for it regardless of how expensive it is. It’s how the economy works.
There is no free lunch. Even if a company pays for it instead of you, their LLM costs per developer will be factored in to what they are willing to provide as compensation. So one way or another, the end result is you get paid for less for the same amount of work today.
Well bucko it’s time to open your wallet. There’s creatives out there who spend at least $1000/month in subscriptions for tools, but without those tools they could never do most of the work they do. And some who buy physical gear like photographs and videographers pay even way more than that for equipment.
Soon it will be the same for developers. Developers really are a spoiled bunch when it comes to paying for things, many will balk at paying $99/year just to publish stuff on an App Store. Everyone just wants free open source stuff. As expectations on developer productivity rises, you will be paying for these AI tools no matter how expensive they get, or you will just be gentrified out of the industry. The choice is yours.
That “if” doesn’t apply to all of us, though. Not everybody is paid by the hour. I’d love to try something like Claude code, but $100 per month is way too expensive for me, and it probably wouldn’t even give me a single extra dollar of income. I think I’ll just wait for the time when local LLMs will be good enough to be a viable alternative.
By the time you can run good enough local LLMs without splurging on sufficiently powerful hardware, those LLMs will look like toys compared to whatever cloud based LLMs are available.
That's a great question. Probably not. IDK. I'm also only paying this much to maintain momentum on a personal project. I also know in a year, these LLM products will change drastically, pricing tiers will transform, etc.. So I can't predict what will happen in a year but things will probably be cheaper.
Edit: On the other hand, the state of the art tools will also be much better in a year, so might keep that high price point!
Am I rationalizing my purchase? Possibly. If I'm not using it daily, I will cancel it, I promise :)
I think there is definitely room to price AI tools way higher. Developers are being slowly boiled like frogs right now. Getting addicted to AI tools to the point they can’t work without them, that’s when you raise the price.
I see it as an investment into my future. I was able to make progress on a personal project with Claude Code which I failed at using other tools. Yes, I will, and apparently have, paid multiple hundreds of dollars to get the project release ready. But I definitely need to keep in mind that I'm not going to at that velocity all the time, which would make the $200 price point not justifiable long term.
The current max pricing is actually as transparent as it has ever been: It's 20% more to use Max than the APIs directly. I am not sure if your feedback is outdated/based on a previous version of reality?
I have a workflow that also uses micro commits.
I keep my older JetBrains IDE open at the same time.
Using feature branches liberally, any successful interaction between me and the LLM in Cursor results in a micro commit. I use the Cursor AI ‘generate commit message’ for speed.
Every so often, I switch over to Jetbrains to use Git Interactive Rebase to tidy up the commits, as the diff viewer is unsurpassed. Then those micro commits get renamed, reordered, squash merged as required. All possible from Git CLI of course, but the Jetbrains Git experience is fantastic IMHO. All their free community edition IDEs have this.
Have they managed to fix the bug where applying changes to a file moves the file to the first in your list of open files? And even pinning it if you have other files pinned? Overall, while I've liked using Cursor, it has many bugs like this that I haven't experienced in other VSCode forks and makes me wonder what they consider to be 1.0 quality.
I keep getting back and forth between Cursor and Zed, but Cursor autocomplete and next cursor prediction are still the best in class between all the competitors, I don't use chats and agents, yet I feel very productive and fast. I sometimes go back to Copilot too just to see how is it going but it has been very delusional so far regarding code suggestions. The only thing I hate about Cursor is the overwrite of some of the shortcuts of vscode, I remapped some and learned some new, and that vim mode plugin is a bit buggy. This and the fact that performance compared with Zed is shit and that's why I go back to Zed sometimes, I'm pondering the idea of just using both and keep them both open
> I'm pondering the idea of just using both and keep them both open
Do it. I've started editing with Zed and just keeping Cursor/Intellij open on the side. (Cursor b/c of the the free student plan, Intellij for school assignments).
I feel spoiled by the performance, especially on promotion displays. I've started noticing some dropped frames in Cursor and measured an avg of 45-60 fps in Intellij (which is somewhat expected for such a huge IDE). I basically exclusively write in Zed, and do everything else in their respective apps.
I think not using agent chat is kind of a missing forest for the trees sort of thing.
That said, I do continue to think that agents are in this weird zone where it's more natural to want to interact through ticketing layer, but you kind of want to editor layer for the final 5%.
Missing forest by chatting about a problem? I keep going back to agents from time to time, never worked for me, and I already spend a lot of time understanding boss tickets(via extra meetings and feedback loops) I don't want explain again down the line, my strong opinion is that I'm paid for doing real work augmented by AI, not assigning "tickets" to the AI
I don't really see why people still use Cursor over tools like Cline / Roo Code. I'm guessing it's as they clearly have a larger viral marketing department than engineers, as the application itself doesn't perform nearly as well, requires you to have another IDE installed and their subscriptions nerf the models context sizes etc...
Because at some point we have to stop riding the treadmill and just pick a tool and use it to make stuff. Cursor was the first to really arrive at a useful agent mode. It does everything I need it to and more. It's not worth it to me to keep hopping to new tools every time a new one becomes the hyped up hot thing.
Like it or not, we're hitting the slope of enlightenment and some of us are ready to be done with the churn for a while.
Cline was agent based from day one and doesn't try to do copilot style tab completes at all. It's been our go to agentic coding app across the majority of our large clients since mid-late 2024. Cursor has been trying to play catch up but has not delivered us the same results.
> doesn't try to do copilot style tab completes at all
Which is another reason why I'll stick with Cursor. Cursor's tab complete can barely be described as Copilot-style, it's nearly a different paradigm and it's what actually got me to pay for it in the first place. I only tried agent mode because it was included with the bundle.
> from day one
July 5, 2024 if going by initial commit. So, yes, technically before Cursor, but Cursor was building agent mode before Cline drew any attention in the mainstream. Cline's first interest on HN dates back to January.
I'll concede that it appears Cline did get to agents first, but it's still a new development in terms of actually drawing interest.
Remember, not everyone is a pro dev. I'm certainly not: and in that context I find Cursor to be incredibly simple and useful.
It's just like VSC, which I was using, but it has these magical abilities. I was using it within a minute of downloading it. Unlike Cline, I guess, whatever that is.
Because my company decided to pay for everyone's cursor and I don't have the bandwidth to spend my time constantly evaluating what's better and pitching it?
I use cursor because it’s way cheaper to pay their monthly subscription than bringing my own key. I’ve tried all tools and in the end the most cost effective one ended up being cursor. In others I’d end up burning $10 a day.
I couldn't possibly disagree with you more that Cline is better than Cursor. Cursor's success isn't because of "a larger viral marketing department"; it's because they made superior software and service.
Cline eats tokens for breakfast, especially for reasoning models. Used it to apply an older patch for a certain React Native lib to a newer version (was not very big but files moved around etc.) and it blew through my free 1 Mio tokens per day for o3 in a few minutes. It worked flawlessly though, but Cursor is just way cheaper.
Because if you’re paying for your own subscription, it’s a way to control costs. If you know how to use it properly, it’s possible to stay within the $20/month spend. Just not if you are tossing trivial tasks to Claude4/GeminiPro and forever topping up tokens.
> their subscriptions nerf the models context sizes etc
You can use the full-context if you prefer that cost/speed tradeoff! Just have to turn on Max Mode.
Cline is great for many users, but a bit of a different product. Lots of Cursor's value come from custom models that run in the background (e.g. Tab, models that gather context, etc.).
A lot of power in social influence. Especially with the younger generations who remix that influence - compound spread of mindshare. Cursor is all over social media.
That's silly. Cursor has the best autocomplete experience, period, and some people prefer that to agent-style interactions.
There's still a ton of low hanging fruit that other Copilot-style autocomplete products don't seem to be picking up, like using clipboard contents, identifying the next place in the file to jump to, etc.
I primarily save time coding with AI with autocomplete, followed by chat, with agentic flows a very distant 3rd, so Cursor is a legitimately better product for me.
It’s not silly at all. There is a lot of hyper activity going in terms of social influence.
I didn’t say cursor has poor UX.
I tab too. And use agent for cheaper work I don’t care too much about. That said, the best autocomplete is arguably evolving and cursor does not own that.
I have too much work and too little time. I tried them all and cursor is the only one with a polished enough experience that made it stick right away. Others might be good but I didn’t have anything near the flowless experience of cursor
Someone said "I don't really see why people still use Cursor over tools like Cline / Roo Code"
And your answer is "A lot of power in social influence.", which is a bit silly when autocomplete is the first form of AI assistance a critical mass of people found intuitive + helpful and Cursor has the best implementation of it... meanwhile Cline/Roo Code don't provide it.
Nah, you definitely don't get it. Some people are here enjoying the act of programming, and Cursor Tab is acting like an improvement on IntelliSense/autocomplete that actually knows what it's doing. Not all of us want to spend half an hour going back and forth with a robot about what it didn't do quite right when we can be in the actual code, tweak a couple lines, and press tab for it to replicate the change in the next 50 now it knows.
Agentic coding is fine, definitely helps me a lot with setup and boilerplate, but finer business logic details and UX changes are now it's strong suit especially if you know WHAT you want but not HOW to explain it in a clear enough format that it can do it without additional prompting.
I actually like programming, and I find typing and having the model autocomplete my changes pretty useful.
I’d rather do that than painstakingly put my request into prose, cross my fingers, and hope the “agent” doesn’t burn 100,000 tokens making stupid and unrelated changes all over my codebase.
I believe in “show, don’t tell,” and autocomplete is the former while agents are the latter.
It is more effective when you have to do a bunch of similar changes. Or when code is standard enough that you just hit tab, and change perhaps one line. Or when parts of code are immediately deduced from context, and substitutions/changes are immediately provided.
I can’t try Background Agents yet because they aren’t available in Privacy Mode. I’m curious if—and how—they’ll roll this out to others, given their guarantee that no code is stored on their servers. According to their security page [1], about 50% of users have Privacy Mode enabled.
I’m also curious how this compares to OpenAI’s Codex. In my experience, running agents locally has worked better for large or complex codebases, especially since setting up the environment correctly can be tricky in those setups.
I am building a jupyter native code agent that interact with jupyter kernel better and can understand data, charts, etc. Not just read/edit jupyter files.
It's an IDE but they don't bother to list on their website what languages you can use the IDE to develop for? I feel like I'm going crazy here. How can they not bother to mention that on their website / marketing?
On VS Code's front page it explicitly lists 12 languages and that it "supports almost every major programming language. Several ship in the box, like JavaScript, TypeScript, CSS, and HTML, but extensions for others can be found in the VS Code Marketplace."
They support every language through their plugin architecture.
"Cursor works with any programming language. We’ve explicitely worked to improve the performance of our custom models — Tab included — on important but less popular languages like Rust, C++, and CUDA."
1.0 is not an upgrade, it's the first stable release. Usually it signifies the arrival of some amount of feature completeness and stability compared to the fast paced 0.x days. Of course semver doesn't really fit neatly most software let alone a user facing GUI application, but socially that's what they're trying to communicate with the 1.0.
These days those versioning is just PR and doesn't mean much like if something is stable. Gmail used to have beta mark for how many years but was still used. Rect native is 0.79 but doesn't mean it's not production ready.
Absolutely hate the scammy dark pattern they introduced in the latest update by hiding the close button on the chat and defaulting the expensive, pay-per-use Max mode to on.
Cursor a lot of respect from our dev team if todays slack messages are anything to go by
bugbot's deep in your PRs now, cleaning up stuff before you even hit review. feels tight. but yeah, claude's off doing its own thing : no plugins, no tabs, just poking at the fs and shell like it owns the box. and and that kinda changes the whole feedback loop. no clue who gets to full LLM-as-OS first, but honestly, both are slowly killing off the old dev setup in their own weird way.
And for months now Cursor on Windows loves to run a 'q' command which doesn't fucking exist..so every time it executes a command line command I get this:
C:\projects\my_project>q^D^C
'q' is not recognized as an internal or external command,
operable program or batch file.
If anyone from Cursor is reading this, we are rolling out MCP server usage analytics where we aggregate (anonymous) usage data across several providers. Would be amazing to include Cursor (reach me at frank@glama.ai). The data will be used to help the community discover the most used (and therefore useful) clients and servers.
I can’t find enough value in this to move me away from Copilot on nvim (with the MCP plugin) and VS Code’s agent mode. Even Zed is better than Cursor at this point, solely because it’s faster. And they all support Claude 4.
I can’t be the only one for which the Python support in Cursor has been absolutely garbage the past week. I’m super disappointed with Cursor. I wanted to love it
There has to be some kind of joke in here about how long it takes people to declare a 1.0 release. "1 million ARR? Not yet. 10 million ARR? Not yet. 300 million ARR? Maybe soon."
I trust that Claude Code is good, and I believe that most people commenting here are truthful to their experiences. However, I have a strange feeling that companies are using bots on these announcements comments.
Maybe I'm being overcautious, but one of the worst things (for me) that came from the AI rush of these past years is this feeling that everything is full of bots. I know that people have preferences, but I feel that I cannot trust anymore that a specific review was really made by a human. I know that this is not something new, but LLMs take it to the next level for me.
Totally agree, but that said, I just fired up Cursor on a paid account and after a few chats immediately hit the same issue I've been facing for weeks:
'Connection failed. If the problem persists, please check your internet connection or VPN'
I've contacted support and they have been no help. You can see tons of people having this issue in user forums. Meanwhile, bypassing the giant monstrosity that is VScode (and then a Cursor as a fork on top of it) gives me no such issues.
So I wouldn't be so dismissive that anyone frustrated with Cursor is a bot.
All of these reviews are irrelevant anyway because of the variations in the problems, skillset, project attributes (size, structure, etc), human variations in prompting, and a million other reasons.
You should just set aside some time to try out different tools and see if you agree there's an improvement.
For trying models, OpenRouter is a big time saver.
Sadly, I don't think this astroturfing is limited to announcement threads. It seems it is becoming increasingly hard to source real human opinions online, even on specialized forums like this or Reddit communities.
I hope that I am wrong, but, if I am not, then these companies are doing real and substantial damage to the internet. The loss of trust will be very hard to undo.
I'm not sure about bots but it looks like they have real peoples on payroll or who paid per comment or something like that. And they trying push narrative 'use it now or you will be left behind' on every place where someone could share experience of using ai tools.
As someone who has two annual subs of Cursor Pro (one from student account and another from Lenny's newsletter), I just spent $100 on Claude Code and I haven't touched Cursor AI for any coding tasks since. If you already spend anything near or over $100 on Cursor, it's no brainer. The agent experience is night and day. No more wrong tool-calling, premature ending of conversation, failure to apply changes or overwriting a whole file with the update snippet. I'm considering upgrading to $200 Claude Max next month for more concurrent sessions. If anyone reading this thinking this is a paid comment, go search for other users' feedback. Claude Code is that good.
As a long-time Cursor user, I just tried Claude Code for the first time two days ago and I found it:
- Burning tokens with constant incorrect command-line calls to read lines (which it eventually gets right but seemingly needs to self-correct 3+ times for most read calls)
- Writing the string "EOF" to the end of the file it's appending to with cat
- Writing "\!=" instead of "!="
- Charged me $7 to write like 23 lines (admittedly my fault since I forgot I kept "/model opus" on)
Minus the bizarre invalid characters I have to erase, the code in the final output was always correct, but definitely not impressive since I've never seen Cursor do things like that.
Otherwise, the agent behavior basically seems the same as Cursor's agent mode, to me.
I know the $7 for a single function thing would be resolved if I buy the $100/month flat fee plan, but I'm really not sure if I want to.
I learned about it from this thread and will buy a month's worth to keep playing with it. (48 hours ago the documentation said it was only supported for Max.)
Hello,
Your Pro plan just got way more powerful with three major upgrades previously available only to Max, Team, and Enterprise users.
Claude Code is now included
Claude Code is a command line tool that gives you direct access to Claude in your >terminal, letting you delegate complex coding tasks while maintaining full control. You can now use Claude Code at no extra cost with your Pro subscription.
> Burning tokens with constant incorrect command-line calls to read lines
See this sort of asinine behavior with cursor too sometimes although it's less grating when you're not being directly billed by the failed command line attempt. Also it's in a text editor it fully controls why is it messing around in the command line to read parts of files or put edits in, seems to be a weird failure state it gets into a few times a project for me.
Yes, this is a paid comment, in the sense that it's probably a bot. 22 day old account, with 1 post, praising Claude.
For more than a year, Anthropic has engaged in an extensive guerrilla marketing effort on Reddit and similar developer-oriented platforms, aiming to persuade users that Claude significantly outperforms competitors in programming tasks, even though nearly all benchmarks indicate otherwise.
> Please don't post insinuations about astroturfing, shilling, brigading, foreign agents, and the like. It degrades discussion and is usually mistaken. If you're worried about abuse, email hn@ycombinator.com and we'll look at the data. [0]
I believe that's talking about general insinuations, e.g. "you're just a russiabot" and the like.
The GP account above with only one comment that is singing the praises of a particular product is obviously fake. They even let the account age a bit so that wouldn't show up as a green account.
The most alarming to me thing is that it seems to be happening at scale. This is one of dozens similar posts I've seen all over the programming communities with similar characteristics (high praise, new-ish accounts, little if any other activity).
Well, I’m not a paid comment, and I agree 100% with the op, and have the exact same experience. I haven’t touched Cursor since paying for Claude Code (max or whatever the $100/mo plan is). That said, I never found Cursor very useful. Claude Code was useful out of the gate, so my experience may not be typical.
I recently wrote a 5+ page internal guide on how I do vibe coding and one of the first sections is cost
Keep in mind much of the guide is about how to move from 30s chats to doing concurrent 20min+ runs
----
Spending
Claude Code $$$ - Max Plan FTW
TL;DR: Start with Claude Max Pro at $100/mo.
I was about $70/day starting day 2 via the pay-as-you-go plan. I bought in $25 increments to help pace. The Max Plan ($100/mo) became attractive around day 2-3, and on week 2 I shifted to $200/mo.
Annoyingly, you have to make the plan decision during your first login to Claude Code, which is confusing as I wanted to trial on pay-as-you-go. (That was a mistake: do Max Pro.) The upgrade flow is pretty broken from this perspective.
The Max Plan at the $100/mo level has a cooldown of 22 question / 5 hour: That does go by fast when your questions are small and get interrupted, or you get good at multitasking. By the time you are serious, the $200/mo is fine.
Other vibe IDEs & LLM providers $$$
I did anywhere from about 50K to 200K tokens a day on Claude 3.7 Sonnet during week 1 on pay-as-you-go, with about a ratio of 300:1 of tokens in:out. Max Plan does not report usage, but for periods I am using it, I expect my token counts to now be higher as I have gotten much better at doing long runs.
The equivalent in OpenAI of using gp4-4o and o3 would be $5-40/day on pay-as-you-go, which seems cheaper for using frontier models… until Max Pro gets factored in.
Capping costs
Not worrying about overages is typically liberating. Max Pro helps a lot here. One of my next experiments is seeing about self-hosting of reasoning models for other AI IDEs. Max Pro goes far, but to do automation and autonomy, and bigger jobs, you need more power.
I was spending around $800 in Cursor and I’ve switched to Claude Code with a $200 subscription and I couldn’t be happier. The experience is way better (although tbh Claude Code is missing some critical features like being able to rollback changes (or, as Cursor calls them “checkpoints”)) but for 99% of my “Vibe Coding”, it’s just great. I usually run 2 to 4 parallel sessions using git worktrees and the speed is absolutely crazy.
Of course not everything is perfect and I still have to check most of the code but if you create a good enough set of “memories” (Claude Code’s version of .cursorrules) it gets stuff right almost all the time.
That sum is ridiculous, but not everyone is only programming for bigcorp. I've had Cursor Pro for something like 3 years on my own dime. So yes hundreds of hundreds of dollars.
How do you own Cursor for 3 years, when even ChatGPT is not that old? The earliest Cursor submission to HN was on October 15, 2023 --- not even 2 years old [0].
Pro since Aug '23 based on my invoices. Sorry, dang can update my wildly wrong timeline if wished :). Not sure when my account was created but feels like forever.
Copilot at $10/mo. lets you use Sonnet 4 in VS Code, which has been working very well for me in agent mode. Curious what Claude Max offers that sets it apart.
It can use any command line tool very well. I just told him "Look up the status of the created systemd servive". It ssh-d to the machine, run "systemctl status", read the output and fixed issues based on that! That was totally unexpected.
I hope this question doesn't sound snarky, it's a legitimate concern that I want to address for myself: how do you ensure that once it ssh's to the machine, it does not execute potentially damaging commands?
Claude code asks you permissions for every command. It also gives you the possibility of marking commands as safe so next time it can use them without asking .
So these agents that people are so excited about spawning in parallel stop and ask you before executing each command they choose to execute? What kind of life is that. I'd rather do something myself than tell 5 AI agents what I want and then keep approving each command they are going to run.
I'm not saying it is better if they run commands without my approval. This whole thing is just doesn't seem as exciting as other people make it out to be. Maybe I am missing something.
It can literally be a single command to ssh into that machine and check if the systemd service is running. If it is in your history, you'd use ctrl+r to lookback anyway. It sounds so much worse asking some AI agent to look up the status of that service we deployed earlier. And then approve its commands on top of that.
I think it's something you have to try in order to understand.
Running commands one by one and getting permission may sound tedious. But for me, it maps closely to what I do as a developer: check out a repository, read its documentation, look at the code, create a branch, make a set of changes, write a test, test, iterate, check in.
Each of those steps is done with LLM superpowers: the right git commands, rapid review of codebase and documentation, language specific code changes, good test methodology, etc.
And if any of those steps go off the rails, you can provide guidance or revert (if you are careful).
It isn't perfect by any means. CC needs guidance. But it is, for me, so much better than auto-complete style systems that try to guess what I am going to code. Frankly, that really annoys me, especially once you've seen a different model of interaction.
I do not think that is a good thing in the long run. More people in fields they know absolutely nothing about? That does not sound like a good thing to me. I am going to become a chemical engineer (something I know absolutely nothing about) or some shit and have an LLM with me doing my job for me. Sounds good I guess?
He has a point, that's quite depressing that a work you had to think and act in order to solve hard problems now became almost the same as scanning barcodes in any supermarket, and it's outright sad that most people are happy about it and being snarky towards anyone that points the hardships that come with it.
Philosophically speaking (not practically) it's like living the industrial revolution again. It's lit! But it's also terrifying and saddening.
Personally it makes me want to savor each day as the world would never be the same again.
I mean most software engineering jobs are not especially exciting. I have done web dev for smaller companies that never had more than a few hundred concurrent users. It is boring CRUD apps all day every day.
Still at least you could have a bit of fun with the technical challenges. Now with AI it becomes completely mind numbing.
I'm with you on this. I'm pouring one out for human skill because I think our ability to do a lot of creative work (coding included) is on the brink of extinction. But I definitely think these are the future
The interesting part of my job is unchanged. Thinking through the design, UX, architecture, code structure, etc were always where I found the fun / challenge. Typing was never the part I was overly fond of.
I mostly use it to understand and fix my bugs/Rust compilation problems that I can't be bothered to fix. 95% happy with the results so far. For coding I use Claude in chat though, as my thoughts are mostly not clear enough at the start to finish a component to my liking. Fixing bugs is easier, though I had to tell it to "not remove features" sometimes. Feature gone, bug gone. ;)
Claude code now automatically integrates into my ide for diff preview. It's not sugar, but it's very low friction, even from the cli.
The system prompt, agent landscape and fundamental behavior is different. Its just like using chatgpt vs openai api. A single chatgpt conversation can go forever because it's not just doing one call for each message you send.
Does anyone know how Claude code compares to using Aider with anthropic API?
I have been using the Claude.ai interface in the past and have switched to Aider with Anthropic API. I really liked Claude.ai but using Aider is a much better dev experience. Is Claude Code even better?
CC is more autonomous, which can be a double edged sword. In big codebases you usually don't want to make large changes and edit multiple files. And even if you do, letting the LLM decide what files to edit increases the chance for errors. I like Aider better aswell. It's a precision tool and with some /run it is pretty flexible for debugging.
Claude Code is much more aggressive at doing stuff than aider is (with sonnet and gemini) in both good and bad ways. You can tell Claude to do a thing and it might churn for many minutes trying to achieve it, while aider is much more likely to do a lot less work then come back to me. Aider feels more like a small sharp tool vs Claude Code as a bulldozer.
They both can just use api credits so I’d suggest spending a few dollars trying both to see which you like.
Claude Code (and Cursor, for that matter) don't commit to git. Fundamentally that's just bonkers to me. Aider does, so each prompt can be /undo'd. I had a chance to use Cursor at work, and if that's how people are interacting with LLMs, it's no wonder we can't agree on this whole "are LLMs useful for programming" thing.
ChatGPT Codex is on another level for agentic workflow though. It's been released to (some?) "plus" ($20/month) subscribers. I could do the same thing manually by making a new terminal, making a new git worktree, and firing up another copy of aider, but the way codex does it is so smooth.
I haven't used Jujutsu / jj much at all. But it seems like a great match to Aider. I wonder how the surrounding dev tooling ecosystem changes as agentic coders become more popular.
What about the Cursor tab? You can't get that from Claude code which is terminal based. For small changes tab is much faster than asking the agent + it's free.
Same experience here. I got a Cursor Pro sub towards the end of 2024. Once Claude Code because available, my Cursor usage dropped _dramatically_ as I got up to speed on how Claude Code worked.
I still prefer Cursor for some things - namely UI updates or quick fixes and explanations. For everything else Claude Code is superior.
Honestly, once you try aider, no other AI coding tool can reach that level of productivity.
What's the best about it, it's open source, costs nothing, and is much more flexible than any other tools. You can use any model you want, either combine different models from different vendors for different tasks.
Currently, I use it with deepseek-r1-0528 for /architect and deepseek-v3-0325 for /code mode. It's better than Claude Code, and costs only a fragment of it.
Once something, like in this case AI, becomes a commodity, open source beats every competition.
I used to use Cursor and just deal with the slow requests for most of the month because it was the most affordable way to leverage an agent for coding, but I didn't find it so much better than Cline or Roo. When I first tried Claude Code, it was immediately clear to me that it worked better, both as an agent and for me, but it was way too expensive. Now with the $200/mo. Max plan, I couldn't be happier.
That said, I still approach it with the assumption that Claude Code is just mashing its fists on the keyboard and that there needs to be really strong, in-loop verification to keep it in line.
Well, sure. Except that in most of the world outside SV, $200/month is expensive.
At least Cursor is affordable to any developer. Because most of the time, even if it’s totally normal, companies act like they’re doing you a favor when they pay for your IDE so most people aren’t going to ask an AI subscription anytime soon.
Even in a European with lower wages compared to US, total cost of a developer will be minimum 5000 euros/per month. And that's just salary with all taxes, not accounting laptop costs, office space, etc.
You just need a 4% increase of productivity to make those $200 worth it.
> Even in a European with lower wages compared to US, total cost of a developer will be minimum 5000 euros/per month. And that's just salary with all taxes, not accounting laptop costs, office space, etc.
lolololol
> You just need a 4% increase of productivity to make those $200 worth it.
who “needs” that and who pays for it?
the employer for both?
high school economics class is not how the world works, regrettably.
I guess the world doesn't work like that because employers don't even understand high school economics.
They'd rather have an employee spend 2 weeks on a task than shell out a few bucks at it, because they don't realize the 2 weeks of salary is more expensive than the external expense.
I was almost exclusively using Claude Code for a couple of weeks, and after recently trying Cursor with Sonnet 4 in MAX mode, I think it now comes close. Those are requests paid on top of the sub price though.
Claude Code is always quite slow for me. I'm on Windows though so I'm not sure if the performance hit comes from WSL. Anyone experienced differences between Windows/Mac/Linux?
Claude Code has been pretty poky for me running under WSL 2 on my Server 2022 box (admittedly, with pretty outdated hardware). It would routinely hang after a lot of use and performance would degrade over time.
Since the last couple of updates I don't seem to have those problems as prominently any more. Plus it seems to have greatly improved its context handling as well -- I've encountered far fewer occurrences where I've had to compact manually.
It's slow between each UI interaction compared to auto-complete type systems. But CC can do much more per step. That's especially true if you use it to write scripts that encapsulate multiple steps of a process. Then it can run the scripts.
Perhaps not coincidentally, that's what efficient (or "lazy", you choose) developers do as well.
I just recently got $200 sub for Claude and it really is worth it. I work with very large codebases and to be honest Cursor is horrible at those. Claude takes time but in the end it can explain how things work in detail, unlike Cursor.
In a similar position – I get Cursor Pro on a free student plan, which means a lot on a student budget. It looks like Anthropic can afford not to offer a cheaper education plan directed to students do so for now.
Recently canceled cursor. I think there's a shift happening right now with the improvements in the ability to process large context sizes and stay on task:
This is basically a gradient of AI output sizes. Initially, with the ability to generate small snippets (autocomplete), and moving up to larger and larger edits across the codebase.
Cursor represents the initial step of AI-assisted traditional coding... but agent mode is reliable now, and can be directed fairly consistently to produce decent output, even in monorepos (IME). Once the output is produced by the agent, Ive found I prefer minimal to no AI for refining it and cleaning it up.
The development techniques are different. In agent mode, there's far more focus on structuring the project, context, and prompts.. which doesn't happen as much in the ai-autocomplete development flow. Once this process shift happened in my workflow, the autocomplete became virtually unused.
So I think this shift toward larger outputs favors agent-focused tools like CC, Aider, Cline, and RooCode (my personal favorite).. over more traditional interfaces with ai-assistance.
Yes, my initial distrust lead me to believe that I should only trust whatever ai agent it is to handle the granular issues in code. But they're fundamentally bad at it unless you spend more time prompting and re-prompting than a human with the right context would spend doing it properly.
Now I've changed my technical planning phase to write in a prompt-friendly way, so I can get AI to bootstrap, structure, boilerplate and usually also do the database setup and service layer, so I can jump right into actually writing the granular logic.
It doesn't save me planning or logic overhead, but it does give me far more momentum at the start of a project, which is a massive win.
If they're small corrections, I generally agree that manual changes are the easier solution. I have been recently trying to correct it and then having it generate a Cursor rule to tell itself to avoid that initial mistake (style and structure situations) in the future. Doesn't always work out, but it's handy when it does.
Cursor had been atrocious. Building on top of an already crappy IDE, you’d hope that they are at least keeping up with VSCode improvements and updates. But they are far behind and instead keep slapping on more garbage.
The agent stuff is largely useless. The tab prediction go nuts every few seconds completely disrupting flow.
I agree it has good completes, but it also hallucinates completes for blank lines or in the middle of typing a one word change and that can be irritating. Still better than plain VS code autocomplete though.
> The tab prediction go nuts every few seconds completely disrupting flow.
This is my main gripe with it, too. It's still been semi-useful at least for some analysis and examination of our code-base, but editing and autocomplete I've not found super useful yet.
If you have a decent GPU or a modern Mac, you can run something like LM Studio (https://lmstudio.ai/) or Ollama (https://ollama.com/), configure the tool such as Cursor to use those models you've downloaded (personally I use Zed.dev), and then everything happens straight on your computer. Responses will be somewhat slower and not as good as state-of-the-art models, but they still can be helpful.
No, he basically means thay companies will not allow LLMs on their own code, I think.
I work in a multinational conglomerate, and we got AI allowed ... 2-3 weeks ago. Before that it was basically banned unless you had gotten permission. We did have another gpt4 based AI in the browser available for a few months before that as well.
I could probably go much lower, and find a model that is dirt cheap but takes a while; but right now the cutting edge (for my own work) is Claude 4 (non-max / non-thinking). To me it feels like Cursor must be hemorrhaging money. The thing that works for me is that I am able to justify those costs working on my own services, that has some customers, and each added feature gives me almost immediate return on investment. But to me it feels like the current rates that cursor charges are not rooted in reality.
Quickly checking Cursor for the past 4 day period:
Requests: 1049
Lines of Agent Edits: 301k
Tabs accepted: 84
Personally, I have very little complaints or issues with cursor. Only a growing wish list of more features and functionality. Like how cool would it be if asynchronous requests would work? Rather than just waiting for a single request to complete on 10 files, why can't it work on those 10 files in parralel at the same time? Because now so much time is spend waiting for the request to complete (while I work on another part of the app in a different workspace with Cursor).
They don't make any money. They are burning VC money. Anthropic and OpenAI are probably also not making moeny, but Cursor is making "more no money" than others.
This is based on what I've read here: https://www.wheresyoured.at/openai-is-a-systemic-risk-to-the... (big AI bear, for what it's worth)
edit: Well, if they shed the other expenses that only really make sense when training future models (research, more data, fewer employees ..) they would be pretty close to break even.
It's like a horse race.
But yeah enjoy the subsidies. It's like the cheap Ubers of yesteryear.
The more interesting, novel, and useful work you wrap the LLM in the more defensible your pricing will be.
That said I think this can describe a lot of agentic code tools - the entire point is that you're not just talking to the raw LLM itself, you're being intermediated by a bunch of useful things that are non-trivial.
I see this with Anthropic most - they seem to have multiple arms in multiple lines of business that go up the abstraction ladder - Claude Code is just one of them. They seem to also be in the customer service automation business as well.
[edit] I think a general trend we're going to see is that "pure" LLM providers are going to try to go up the abstraction ladder as just generating tokens proves unprofitable, colliding immediately with their own customers. There's going to be a LOT of Sherlocking, and the LLM providers are going to have a home field advantage (paying less for inference, existing capability to fine-tune and retrain, and looooooots of VC funding).
However they will benefit in a way like they benefit from faster server processors: they still have competition and need to fight to stay relevant.
The customers take a lot of the value (which is good).
While there is a lot of fear around AI and it's founded I do love how no one can really dominate it. And it has Google (new new IBM) on it's toes.
They do need to develop sustainable end-user products, or be purchased by larger players, or liquidate.
It's hard to add sophisticated abstractions though, because they are all selling text by the pounds (kilos?). So it feels the same as vendor lock for a cucumber seller, doesn't it? The seller can sell you an experience that would lock you in, but aside from it there is no moat since anyone can sell cucumbers.
It doesn't take much more of a stretch to imagine teams of agents, coordinated by a "programme manager" agent, with "QA agents" working to defined quality metrics, "architect" agents that take initial requirements and break them down into system designs and github issues, and of course the super important "product owner" agent who talks to actual humans and writes initial requirements. Such a "software team system" would be another abstraction level above individual agents like Codex.
As native LLM task completion horizons increase another order of magnitude, so much of this falls out of the window.
Inference cost is plummeting. It’s like the cheap Ubers of yesteryear, if the cost of hiring a driver dropped by a factor of a thousand in the past three years.
> The cost of LLM inference has dropped by a factor of 1,000 in 3 years.
— https://a16z.com/llmflation-llm-inference-cost/
AI startups are not profitable because they are throwing vast sums of money at growth and R&D, not because inference is unaffordable.
You are asking if infinite usage is affordable.
Pet food is not subsidized in my country nor the EU. If any countries do subsidize pet food, they are the exception. Maybe the US? Pet food is often manufactured from the waste of other processes, including the human food industry, but that is not a subsidiary.
<rant>
Think of it like this: imagine if lactose-free or gluten-free food could be bought only with a prescription. Sadly the prices are already high as it is for gluten-free, but I would rather not get into the reasons here. :)
My girlfriend (LA, US) just left 1k USD on 2 visits to the vet with her cat, for some cheap ass antibiotics, and a "specific prescription-only food". Crazy. All that would have been "free" (not quite, but you know) or at a very low cost for humans around here, in Europe. Not the gluten-free food though!
</rant>
This also goes to a personal issue that why would you feed your pet a waste product. My dog gets food I cook for him just like myself. There are tons of crock pot recipes online for safe cheap high quality dog food.
Switching costs are zero and software folks are keen to try new things.
That makes the opportunity cost of switching significant.
(I'm not really a coder/programmer/engineer).
You can open up to three parallel chat tabs by pressing Cmd+T
https://docs.cursor.com/kbd
Each chat tab is a full Agent by itself!
The problem with generative ai workloads: The costs rise linerly with the number of requests, because you need to compute every query.
Both are genuine questions.
I have used or rather use Claud with CoPilot and I find it pretty useful but at times it gets stuck in niche areas.
What does this measurement mean?
1049 / (4 * 8) ~= 32 seconds, on average. Doesn't look like much waiting to me.
Also another issue I am starting to see is the lack of shared MCP servers. If I have VSCode, Cursor, and Claude open, each one is running its own instance of the MCP server. You can imagine that with a dozen or so MCP's, the memory footprint becomes quite large for no reason.
I don’t think the future of agentic software development is in an IDE. Claude Code gives me power to orchestrate - the UX has nothing to do with terminal; it just turns out an agent that lives on the OS and in the filesystem is a powerful thing.
Anthropic can and will evolve Claude Code at a pace cursor cannot evolve IDE abstractions. And then yea - they are designing the perfect wrapper because they are also designing the model.
Long bet is Claude Code becomes more of an OS.
It… sure is something. I’m still thinking about if it’s horrible, visionary, or both.
> Automate Content: Like this very post. I use Wispr Flow to talk with Claude, explain the topic and tell it to read my past blog posts to write in my style.
This is why I think the future may be less about “agents” and more about “intelligent infrastructure”
I don’t want to chat. I want to point a wand and make things happen.
What I did do, because my personal projects aren't too complex, is moved the model from Opus to Sonnet which is about 1/5 the cost.
For day-to-day stuff I have ProxyAI (on IntelliJ, continue.dev works for this too) pointed at Mistral's Codestra for auto-complete and to Claude 4 for chat.
Claude Code is just for giving the bot tasks to do without me being deeply involved in the work.
(edit) I just saw that pro is getting a rate-limited option for Claude code for the sonnet model only. I haven't tried it out but will give it a go sometime. https://support.anthropic.com/en/articles/11145838-using-cla...
The core suggestion is to point specifically at the files you want it to read and use as a reference, otherwise it might go read some huge file for no reason. Also the tokens used depend on the size of the project.
Generally, if I'm doing something I can box, I'd use chatgpt and copy it in myself. If I want something matching the style, I'd use a guided Roo Code.
Besides, if you want something inexpensive, using Gemini 2.0 Flash as a backend is completely free. Google provides an API key at no cost.
Didn't try anything agentic within Emacs yet, don't find that helpful enough so far.
As of last week you can insert a link to a plain-text file in a chat buffer to include its contents in the prompt. It must be on a line by itself. In Markdown it looks
[like this](/path/to/file)
with Org links in Org chat buffers.
This feature is disabled by default to minimize confusion. To enable it you can flip the header line button that says "ignoring media" to "sending media". This works for sending images and other media too, if the model supports it.
I have a global bind for gptel-send (C-c g).
Then, in any buffer, I typically type C-u C-c g.
This lets me customize the prompt and lots of parameters, such as the context, before gptel-send is actually called.
My only wish is that Cursor had partnered with Zed. vscode isn't enjoyable.
The $200 plan so far has been fine. Only had it once that it looked like I might hit the limit soon, but that was a very heavy refactoring task
Why is that?
If it is delivering that value indeed, then 100-200 dollars each month is exactly what that professional SWE is worth.
That SWE is going to have to pivot into something adjacent to provide value-add, and if the only value-add is "I can review and approve/reject", then it's only a matter of time[1] that there are no SWEs who can review and then approve/reject.
And that is assuming that the LLM coding performance remains where it is now, and does not improve to the review, orchestration, design and planning levels.
-------------------------
[1] Due to attrition
Unless you do nothing else with your time I'm not sure how you'd utilize the $100/mo plan fully.
I admit their transparency around limits is a bit questionable. Why don't they just list out the tokens and remaining time?
Sometimes I'll just experiment with weird stuff and end up reverting it or delete it afterword's. Also fun to build really nice data visualizations for data I'm working with.
> Pro ($20/month): Average users can send approximately 45 messages with Claude every 5 hours, OR send approximately 10-40 prompts with Claude Code every 5 hours. [0]
I can probably run through that in 5 minutes.
[0] https://support.anthropic.com/en/articles/11145838-using-cla...
[1]: https://gofastmcp.com
Cursor is essentially only the wrapper for running agents. I still do my heavy lifting in Jetbrains products.
It actually works out well because I can let Cursor iterate on a task while I review/tweak code.
Using Windsurf plugins in JB ides has been working for me, albeit not as powerful yet as the Windsurf VS Code fork.
Last time I tried, it didn't support RubyMine (whomp)
Knowing what tools are better for what really helps.
https://github.com/tuananh/hyper-mcp
it's a MCP server with WASM plugin system, packaged, signed & published via OCI registry.
The one issue I've run into is that the VSCode version Cursor uses is several months old, so we're stuck using older extensions until they update.
[1] https://open-vsx.org/
``` { "extensionsGallery": { "serviceUrl": "https://marketplace.visualstudio.com/_apis/public/gallery", "itemUrl": "https://marketplace.visualstudio.com/items", "cacheUrl": "https://vscode.blob.core.windows.net/gallery/index", } } ```
https://github.com/getcursor/cursor/issues/2195
This is technically against the ToS of the extension gallery though
We recently launched Zapier MCP, we host the servers for you: https://zapier.com/mcp
i wrote a MCP with plugin system where you only need to run 1 instance and add plugins via config file.
https://github.com/tuananh/hyper-mcp
This is good user feedback. If Cursor is "Claude + VSCode", why do you need the other 2 open?
Autocomplete: yes
Asking questions: yes
I know everybody is all into agent mode and it isn't a Cursor specific thing, I'm just confused why everybody is into it.
my usecases have been building relatively self-contained, well-specified components of distributed systems out of academic papers and my own design notes / prototype skeletons, in rust. there’s a lot of context for the agent to ground against, and it can blow out ideas into their implications relatively well ime.
the experience of fixing up lots of pesky mistakes that you yourself would never make is kinda annoying.
STDIO MCP is really just a quick hack.
I’m particularly interested in the release of BugBot. The docs mention it looks at diffs but I hope it’s also scanning through the repository and utilizing full context. Requesting copilot to do a review does the same thing but because it’s only looking at diffs the feedback it provides is pretty useless, mainly just things that a linter could catch.
And every time I find it having diverged further from VSCode compatibility.
This wouldn’t be so bad if it was an intentional design choice but it seems more that Microsoft is starting to push them out? Like MS Dev Containers plugin is still recommended by some leftover internal, but if you install it you get pushed to a flow that auto uninstalls it and installs Remote Containers by Anysphere (which works differently and lacks support for some features). And I end up rebuilding my Dev Container once more… I also noticed recent extensions such as the Postgres one from MS also doesn’t exist.
Cursors tab predictions are still a bit better and snappier but I feel like Zed is a better editor experience over all and I don’t rely on AI anyway. Agent mode works pretty well for me though. Also cursor leaks memory pretty bad for me.
There’s still room for improvement but Zed is working on fixes and improvements at a high pace and I’m already pretty happy with where it’s at.
Had similar issues earlier, now it works.
Also it's great that I do not need to use the vscode ecosystem. Zed is snappy, has a great UI and now a good assistant too.
It feels almost as if VSCode is not adding new features and is in maintenance mode for now. I have no idea if that's actually true, but if this continues, a fork will be easily maintainable.
When Ampcode took it all away from me, I found I enjoyed the actual AI-assisted coding much more than configuring. Of course, largely because it just worked. Granted, I had enough experience with other AI tools to manage my expectations.
When reviewing the changes made from agent mode, I don’t know why the model made the change or whether the model even made the change versus a tool call making the change. It’s a pain to go fish out the reason from a long response the model gives in the chat.
Example: I recently asked a model to set up shadcn for a project, but while trying to debug why things looked pretty broken, I had sift through a bunch of changes that looked like nasty hallucinations and separate those from actual command line changes that came from shadcn's CLI the model called. Ended up having to just do things the old fashioned way to set things up, reading the fine manual and using my brain (I almost forgot I had one)
It would be nice if above every line of code, there’s a clear indication of whether it came from a model and why the model made the change. Like a code comment, but without littering the code with actual comments
Hand written code needs to be distinguishable and considered at a higher priority for future code generation context
It’s called git!
Even if you were willing to deal with your agent making commits and having to fiddle with git to break those up or undo them or modify them, it still doesn't solve OP's need. They want to know what bits of diff apply to which thinking tokens. The prompt and the output are necessary, but there's no mapping between the final diff and the stream of tokens that came out of the LLM (especially in a commit). But that information is known by the tooling already! When generating the diff, the LLM had to output code changes. So you can map those changes to where they originated from. It's not just "what did I tell the LLM to get this change" it's "what was the thought process that the LLM went through after getting my prompt to end up with this line change"
where is the splashy overproduced video? where is the promises of agi? where is the "we are just getting started" / "we cant wait to see what you'll build"? how do i know what to think if you aren't going to tell me what to think?
edit: oh haha https://x.com/cursor_ai/status/1930358111677886677
My recommendation to anyone is to use ask the most then tell it to “implement what we discussed” when it looks good.
Hope that helps
In my experience, next edit is a significant net positive.
It fixes my typos and predicts next things I want to do in other lines of the same file.
For example, if I fix a variable scope from a loop, it automatically scans for similar mistakes nearby and suggests. Editing multiple array values is also intuitive. It will also learn and suggest formatting prefences and other things such as API changes.
Sure, sometimes it suggests things I don't want but on average it is productive to me.
It's also somehow tracking what I look at, because I can look up an API and then return to the original source file and the first autocomplete suggestion is exactly what I was looking up, even though there would be no context to suggest it.
It's really quite magical, and a whole different level from Copilot.
That hasn't been my experience with Copilot next edit feature.
It often understands exactly what I'm doing and I'm able to just tab tab around the file. Like my example about variable loop scope fixing.
My experience is that Copilot does everything you said including considering files I viewed previously to better understand what I'm trying to do.
I think this is an artifact of Cursor being a closed-source fork of an open-source project, with a plugin architecture that's heavily reliant on the IDE at least being source-available. And, frankly, taking an open-source project like VsCode and commercializing it without even making it source-available is a dishonorable thing to do, and I'm rooting against them.
Developers use Vim, JetBrains, Emacs, VSCode, and many other tools—what makes you think they’ll switch to your fork?
Its so painful - the model never knows the directory in which it is supposed to be and goes on a wild goose chase of searching in the wrong repo. I have to keep guiding it to the right repo. Anyone here has had success with such a setup?
Keep it short. It's enough for it to realize it needs to navigate directories.
I just have a bunch of markdown files with various prompts that I drop into context when I need them.
Overall, I am having hard time with code autocompletion in IDE. I am using Claude desktop to search for information and bounce off ideas, but having it directly in IDE – I find it too disturbing.
Also there is this whole ordeal with VSCode Marketplace no longer available in Cursor.
I'm not saying AI in IDE is bad, it's just I personally can't get into it to actually feel more productive.
If you sell $1.00 USD for $0.90 you can get nearly unlimited revenue (until you run out of cash).
Also, Trae being $10 for more requests makes Cursor far less appealing to me.
Rust is easier for me than shell scripting so I started writing what I needed and remembered Zed added agent mode. I decided to give it a shot. I had it use Claude 4 with my api tokens.
It wrote the entire program, tested it, debugged it. It made some bad assumptions and I just steered it towards what I needed. By the end of about an hour, I had my complete fix plus an entire ZFS management layer in Rust.
It did cost $11, but that is a drop in the bucket for time saved. I was impressed.
Just sharing this because I got real and measured value recently that is way beyond the widely shared experience.
As an aside: $11 for an hour of Claude 4 seems expensive? I’ve been using Claude 4 (through a Zed Pro subscription, not with my own keys) and I haven’t hit the subscription limits yet. Were you using burn mode, or?
* BYO model or not
* CLI, UI, VSC-plugin or web
* async/sync
* MCP support
* context size
* indexed or live grep-style search
There's probably like 10 more.
Does anyone know if it's GitHub-only or can it be used as a CLI (i.e., Aider replacement)?
Anyway https://github.com/SWE-agent/SWE-agent/blob/v1.1.0/docs/usag... and https://github.com/SWE-agent/SWE-agent/blob/v1.1.0/docs/conf...
Recently, there was a post with detailed evidence suggesting Cursor was intentionally throttling requests [1], including reverse engineering and reproducible behaviors. The team initially responded with a "happy to follow up", but later removed their replies that got downvoted, and banned the OP from posting further updates.
Their response sounded AI-generated too, which wasn't very surprising based on the way they handle customer support [2]. I wish they were more open to criticism instead of only claiming to be transparent.
[1] https://www.reddit.com/r/cursor/comments/1kqj7n3/cursor_inte...
[2] https://news.ycombinator.com/item?id=43700931
Im super happy with it. I’m not sure how it compares to other coding agents though.
It has showed promise, enough to quell my FOMO about other IDEs, since I am extremely happy with the Jetbrains suite otherwise.
Also had a few misses. But in general it is ok. Still prefer ai assistant, because i can then direct the result into a certain direction. It also feels faster, it probably is not because of the manual stuff involved.
I believe it’ll get much better as LLMs start editing code by invoking refactoring tools (rename, change signature, etc.) rather than rewriting the code line by line, as this will let them perform large-scale changes reliably in a way that's similar to how software engineers do them now using IDE tools
Evidently not
It's also missing quite a few features still (like checkpoints, stopping the agent mid-task to redirect, etc). But the core feature set is mostly there.
Luckily, it should be coming with the regular $20 Pro subscription in the near future, so it should be easier to demo and get a feel for it without having to jump in all the way.
Maybe in the US? I will never pay 100$ for a subscription and I despise that people normalized it by even buying this stuff instead of saying "no, that's way too expensive".
Things have a price for a reason. It’s up to you whether it’s worth paying that or not.
I’m trying to convey that if a tool increases your efficiency by more than it costs then it’s worth paying for it regardless of how expensive it is. It’s how the economy works.
Soon it will be the same for developers. Developers really are a spoiled bunch when it comes to paying for things, many will balk at paying $99/year just to publish stuff on an App Store. Everyone just wants free open source stuff. As expectations on developer productivity rises, you will be paying for these AI tools no matter how expensive they get, or you will just be gentrified out of the industry. The choice is yours.
Edit: On the other hand, the state of the art tools will also be much better in a year, so might keep that high price point!
Am I rationalizing my purchase? Possibly. If I'm not using it daily, I will cancel it, I promise :)
You can commit checkpoints prior to each major prompt and use any IDE’s builtin visual diff versus last commit. Then just rebase when the task is done
Do it. I've started editing with Zed and just keeping Cursor/Intellij open on the side. (Cursor b/c of the the free student plan, Intellij for school assignments).
I feel spoiled by the performance, especially on promotion displays. I've started noticing some dropped frames in Cursor and measured an avg of 45-60 fps in Intellij (which is somewhat expected for such a huge IDE). I basically exclusively write in Zed, and do everything else in their respective apps.
That said, I do continue to think that agents are in this weird zone where it's more natural to want to interact through ticketing layer, but you kind of want to editor layer for the final 5%.
Because we're developers with things to build and we don't have time to play with every AI tool backed by the same LLM.
Like it or not, we're hitting the slope of enlightenment and some of us are ready to be done with the churn for a while.
Which is another reason why I'll stick with Cursor. Cursor's tab complete can barely be described as Copilot-style, it's nearly a different paradigm and it's what actually got me to pay for it in the first place. I only tried agent mode because it was included with the bundle.
> from day one
July 5, 2024 if going by initial commit. So, yes, technically before Cursor, but Cursor was building agent mode before Cline drew any attention in the mainstream. Cline's first interest on HN dates back to January.
I'll concede that it appears Cline did get to agents first, but it's still a new development in terms of actually drawing interest.
Cursor Agent works great, too.
Most importantly, everything is $20/month, instead of possibly $20/day with Cline or Roo.
It's just like VSC, which I was using, but it has these magical abilities. I was using it within a minute of downloading it. Unlike Cline, I guess, whatever that is.
You can use the full-context if you prefer that cost/speed tradeoff! Just have to turn on Max Mode.
Cline is great for many users, but a bit of a different product. Lots of Cursor's value come from custom models that run in the background (e.g. Tab, models that gather context, etc.).
There's still a ton of low hanging fruit that other Copilot-style autocomplete products don't seem to be picking up, like using clipboard contents, identifying the next place in the file to jump to, etc.
I primarily save time coding with AI with autocomplete, followed by chat, with agentic flows a very distant 3rd, so Cursor is a legitimately better product for me.
I didn’t say cursor has poor UX.
I tab too. And use agent for cheaper work I don’t care too much about. That said, the best autocomplete is arguably evolving and cursor does not own that.
And your answer is "A lot of power in social influence.", which is a bit silly when autocomplete is the first form of AI assistance a critical mass of people found intuitive + helpful and Cursor has the best implementation of it... meanwhile Cline/Roo Code don't provide it.
Your beloved cursor will go all in on this front, less and less priority on focused cursors in the editor.
Agentic coding is fine, definitely helps me a lot with setup and boilerplate, but finer business logic details and UX changes are now it's strong suit especially if you know WHAT you want but not HOW to explain it in a clear enough format that it can do it without additional prompting.
I’d rather do that than painstakingly put my request into prose, cross my fingers, and hope the “agent” doesn’t burn 100,000 tokens making stupid and unrelated changes all over my codebase.
I believe in “show, don’t tell,” and autocomplete is the former while agents are the latter.
And it's not "my beloved cursor", not sure why you're being such an absolute weirdo about this.
It is more effective when you have to do a bunch of similar changes. Or when code is standard enough that you just hit tab, and change perhaps one line. Or when parts of code are immediately deduced from context, and substitutions/changes are immediately provided.
I’m also curious how this compares to OpenAI’s Codex. In my experience, running agents locally has worked better for large or complex codebases, especially since setting up the environment correctly can be tricky in those setups.
[1] https://www.cursor.com/security#privacy-mode-guarantee
video demo here, https://x.com/ob12er/status/1930439669130637482?s=46&t=2jNrj...
"Cursor works with any programming language. We’ve explicitely worked to improve the performance of our custom models — Tab included — on important but less popular languages like Rust, C++, and CUDA."
Hundreds of languages supported: https://code.visualstudio.com/docs/languages/overview
0 - https://x.com/thisritchie/status/1930598413587959827
Cursor a lot of respect from our dev team if todays slack messages are anything to go by
C:\projects\my_project>q^D^C 'q' is not recognized as an internal or external command, operable program or batch file.
1.0 my ass.
Will be adding the Add to cursor button to Glama later today today.
https://glama.ai/mcp/servers
If anyone from Cursor is reading this, we are rolling out MCP server usage analytics where we aggregate (anonymous) usage data across several providers. Would be amazing to include Cursor (reach me at frank@glama.ai). The data will be used to help the community discover the most used (and therefore useful) clients and servers.
Maybe I'm being overcautious, but one of the worst things (for me) that came from the AI rush of these past years is this feeling that everything is full of bots. I know that people have preferences, but I feel that I cannot trust anymore that a specific review was really made by a human. I know that this is not something new, but LLMs take it to the next level for me.
'Connection failed. If the problem persists, please check your internet connection or VPN'
I've contacted support and they have been no help. You can see tons of people having this issue in user forums. Meanwhile, bypassing the giant monstrosity that is VScode (and then a Cursor as a fork on top of it) gives me no such issues.
So I wouldn't be so dismissive that anyone frustrated with Cursor is a bot.
Not GP, but my suspicions are actually of the other end of the spectrum - i.e., it's the glowing reviews of AI things that make my bot-sense tingle.
Though I usually settle on the idea that they (the reviewers) are using LLMs to write/refine their reviews.
You should just set aside some time to try out different tools and see if you agree there's an improvement.
For trying models, OpenRouter is a big time saver.
I hope that I am wrong, but, if I am not, then these companies are doing real and substantial damage to the internet. The loss of trust will be very hard to undo.
- Burning tokens with constant incorrect command-line calls to read lines (which it eventually gets right but seemingly needs to self-correct 3+ times for most read calls)
- Writing the string "EOF" to the end of the file it's appending to with cat
- Writing "\!=" instead of "!="
- Charged me $7 to write like 23 lines (admittedly my fault since I forgot I kept "/model opus" on)
Minus the bizarre invalid characters I have to erase, the code in the final output was always correct, but definitely not impressive since I've never seen Cursor do things like that.
Otherwise, the agent behavior basically seems the same as Cursor's agent mode, to me.
I know the $7 for a single function thing would be resolved if I buy the $100/month flat fee plan, but I'm really not sure if I want to.
The $100/mo max plan lets you use a Claude Code with a fixed bill. There’s some usage limits though.
See this sort of asinine behavior with cursor too sometimes although it's less grating when you're not being directly billed by the failed command line attempt. Also it's in a text editor it fully controls why is it messing around in the command line to read parts of files or put edits in, seems to be a weird failure state it gets into a few times a project for me.
For more than a year, Anthropic has engaged in an extensive guerrilla marketing effort on Reddit and similar developer-oriented platforms, aiming to persuade users that Claude significantly outperforms competitors in programming tasks, even though nearly all benchmarks indicate otherwise.
[0] https://news.ycombinator.com/newsguidelines.html
The GP account above with only one comment that is singing the praises of a particular product is obviously fake. They even let the account age a bit so that wouldn't show up as a green account.
Keep in mind much of the guide is about how to move from 30s chats to doing concurrent 20min+ runs
----
Spending
Claude Code $$$ - Max Plan FTW
TL;DR: Start with Claude Max Pro at $100/mo.
I was about $70/day starting day 2 via the pay-as-you-go plan. I bought in $25 increments to help pace. The Max Plan ($100/mo) became attractive around day 2-3, and on week 2 I shifted to $200/mo.
Annoyingly, you have to make the plan decision during your first login to Claude Code, which is confusing as I wanted to trial on pay-as-you-go. (That was a mistake: do Max Pro.) The upgrade flow is pretty broken from this perspective.
The Max Plan at the $100/mo level has a cooldown of 22 question / 5 hour: That does go by fast when your questions are small and get interrupted, or you get good at multitasking. By the time you are serious, the $200/mo is fine.
Other vibe IDEs & LLM providers $$$
I did anywhere from about 50K to 200K tokens a day on Claude 3.7 Sonnet during week 1 on pay-as-you-go, with about a ratio of 300:1 of tokens in:out. Max Plan does not report usage, but for periods I am using it, I expect my token counts to now be higher as I have gotten much better at doing long runs.
The equivalent in OpenAI of using gp4-4o and o3 would be $5-40/day on pay-as-you-go, which seems cheaper for using frontier models… until Max Pro gets factored in.
Capping costs
Not worrying about overages is typically liberating. Max Pro helps a lot here. One of my next experiments is seeing about self-hosting of reasoning models for other AI IDEs. Max Pro goes far, but to do automation and autonomy, and bigger jobs, you need more power.
user: christophilus
created: October 16, 2014
karma: 11683
How do you own Cursor for 3 years, when even ChatGPT is not that old? The earliest Cursor submission to HN was on October 15, 2023 --- not even 2 years old [0].
[0] https://news.ycombinator.com/item?id=37888477
And no I'm not a bot but feel as you wish.
What are you doing that costs that much?
I refactored a whole code base in cursor for < $100 (> 200k lines of code).
I don't use completions though. Is that where the costs add up?
You can configure it so that you use your API keys, which means you just pay cost but o3 is expensive
Also, Copilot's paid version is free for developers of popular FOSS projects.
I'm not saying it is better if they run commands without my approval. This whole thing is just doesn't seem as exciting as other people make it out to be. Maybe I am missing something.
It can literally be a single command to ssh into that machine and check if the systemd service is running. If it is in your history, you'd use ctrl+r to lookback anyway. It sounds so much worse asking some AI agent to look up the status of that service we deployed earlier. And then approve its commands on top of that.
Running commands one by one and getting permission may sound tedious. But for me, it maps closely to what I do as a developer: check out a repository, read its documentation, look at the code, create a branch, make a set of changes, write a test, test, iterate, check in.
Each of those steps is done with LLM superpowers: the right git commands, rapid review of codebase and documentation, language specific code changes, good test methodology, etc.
And if any of those steps go off the rails, you can provide guidance or revert (if you are careful).
It isn't perfect by any means. CC needs guidance. But it is, for me, so much better than auto-complete style systems that try to guess what I am going to code. Frankly, that really annoys me, especially once you've seen a different model of interaction.
But a beginner in system administration can also do it fast.
Philosophically speaking (not practically) it's like living the industrial revolution again. It's lit! But it's also terrifying and saddening.
Personally it makes me want to savor each day as the world would never be the same again.
I mean most software engineering jobs are not especially exciting. I have done web dev for smaller companies that never had more than a few hundred concurrent users. It is boring CRUD apps all day every day.
Still at least you could have a bit of fun with the technical challenges. Now with AI it becomes completely mind numbing.
Claude code now automatically integrates into my ide for diff preview. It's not sugar, but it's very low friction, even from the cli.
And as of the latest release, has VSCode/Cursor/Windsurf integration.
I have been using the Claude.ai interface in the past and have switched to Aider with Anthropic API. I really liked Claude.ai but using Aider is a much better dev experience. Is Claude Code even better?
They both can just use api credits so I’d suggest spending a few dollars trying both to see which you like.
ChatGPT Codex is on another level for agentic workflow though. It's been released to (some?) "plus" ($20/month) subscribers. I could do the same thing manually by making a new terminal, making a new git worktree, and firing up another copy of aider, but the way codex does it is so smooth.
I still prefer Cursor for some things - namely UI updates or quick fixes and explanations. For everything else Claude Code is superior.
What's the best about it, it's open source, costs nothing, and is much more flexible than any other tools. You can use any model you want, either combine different models from different vendors for different tasks.
Currently, I use it with deepseek-r1-0528 for /architect and deepseek-v3-0325 for /code mode. It's better than Claude Code, and costs only a fragment of it.
Once something, like in this case AI, becomes a commodity, open source beats every competition.
At least Cursor is affordable to any developer. Because most of the time, even if it’s totally normal, companies act like they’re doing you a favor when they pay for your IDE so most people aren’t going to ask an AI subscription anytime soon.
I mean, it will probably come but not today.
> Except that in most of the world outside SV
You just need a 4% increase of productivity to make those $200 worth it.
lolololol
> You just need a 4% increase of productivity to make those $200 worth it.
who “needs” that and who pays for it?
the employer for both?
high school economics class is not how the world works, regrettably.
They'd rather have an employee spend 2 weeks on a task than shell out a few bucks at it, because they don't realize the 2 weeks of salary is more expensive than the external expense.
Plus development work is quite bursty — a productivity gain for developers does not necessarily translate into more prospects in a sales pipeline.
It's companies asking programmers to use AI, not vice versa.
Since the last couple of updates I don't seem to have those problems as prominently any more. Plus it seems to have greatly improved its context handling as well -- I've encountered far fewer occurrences where I've had to compact manually.
Perhaps not coincidentally, that's what efficient (or "lazy", you choose) developers do as well.
You can generally do map-reduce, also you can have separate git worktrees and have it work on all your tickets at the same time.
am i missing that much ?
Traditional code editing -> autocomplete -> file editing -> agent mode
This is basically a gradient of AI output sizes. Initially, with the ability to generate small snippets (autocomplete), and moving up to larger and larger edits across the codebase.
Cursor represents the initial step of AI-assisted traditional coding... but agent mode is reliable now, and can be directed fairly consistently to produce decent output, even in monorepos (IME). Once the output is produced by the agent, Ive found I prefer minimal to no AI for refining it and cleaning it up.
The development techniques are different. In agent mode, there's far more focus on structuring the project, context, and prompts.. which doesn't happen as much in the ai-autocomplete development flow. Once this process shift happened in my workflow, the autocomplete became virtually unused.
So I think this shift toward larger outputs favors agent-focused tools like CC, Aider, Cline, and RooCode (my personal favorite).. over more traditional interfaces with ai-assistance.
Now I've changed my technical planning phase to write in a prompt-friendly way, so I can get AI to bootstrap, structure, boilerplate and usually also do the database setup and service layer, so I can jump right into actually writing the granular logic.
It doesn't save me planning or logic overhead, but it does give me far more momentum at the start of a project, which is a massive win.
The agent stuff is largely useless. The tab prediction go nuts every few seconds completely disrupting flow.
This is my main gripe with it, too. It's still been semi-useful at least for some analysis and examination of our code-base, but editing and autocomplete I've not found super useful yet.
What about Gitlab instead of GitHub, is there an equivalent to cursor 1.0 product?
Git host doesn't really make a difference.
I work in a multinational conglomerate, and we got AI allowed ... 2-3 weeks ago. Before that it was basically banned unless you had gotten permission. We did have another gpt4 based AI in the browser available for a few months before that as well.