Clarifying our pricing
cursor.com
cursor.com
This news coupled with google raising the new gemini flash cost by 5x, azure dropping their startup credits, and 2-3 others(papers showing RL has also hit a wall for distilling or improving models), are now solid signals that despite what Sam altman says, intelligence will NOT be soon too cheap to meter. I think we are starting to see the squeeze from the big players. Interesting. I wonder how many startups are betting on models becoming 5-10x cheaper for their business models. If on device models don't get good, I bet a lot of them are in big trouble
[1] https://www.investing.com/news/economy-news/anysphere-hires-...
Nvidia sold $35B of just datacenter GPUs last year. Of which the vast majority will be used for AI.
Cerebra entire revenue last year was only $78M. That’s three orders of magnitude smaller than Nvidia datacenter GPU business. Scaling a company 10X in a year is a pretty hard thing to do, and it’s not a question of money, it’s a question of people and organisation. So much stuff in a business breaks when it scales 10X, that it take months to years to fix enough stuff to support another 10x growth spurt without everything just imploding.
The insane thing here is that $35B worth of GPUs will be worth more like $350m in a few years. Or less. Who can keep up with that???
any reference for this?
I’m not convinced that these price increases represent an attempt to squeeze more profit out of a saturated market.
To me they look an awful lot like people realising that the sheer compute cost associated with modern models makes the historical zero-marginal cost model of software impossible. API calls to LLM models have far more in common with making calls to EC2 or Lambda for compute, than they do a standard API calls for SaSS.
A lot of early LLM based business models seemed to assume that the historical near zero-marginal cost of delivery for software would somehow apply to hosted LLM models which clearly isn’t the case.
You mix that in with rising datacenter costs, driven by lack of available electricity infrastructure to meet their demands, plus everyone trying to grab as much LLM land as possible, which requires more datacenters, more faster. And the result is rapidly increasing base costs for compute. Which we’re now seeing reflected in LLM pricing.
For me the thing that stands out about LLMs, is that their compute costs are easily 100-10000x greater per API call than a traditional SaSS API call. That fact alone should be enough for people to realise that the historically bottomless VC money that normal funds this stuff, isn’t quite a bottomless as it needs to be to meaningfully subsidise consumer pricing.
LLM are still to new, and still advancing to quickly for optimisation to take place. It’s like we’re back in the MHz wars of old between CPU manufacturers. The goal is just more performance, regardless of cost, because it was clear that even in the consumer space, people wanted more performance.
Then we hit a kind of plateau in last 10 years, where basic compute is so powerful that your average consumer is not longer upgrading every year for better performance. A 5 year old machine has enough performance for most people. Then the focus on energy efficiency kicked in, because people didn’t want faster computers, they wanted battery life and cheaper computers.
No doubt we’ll see the same with LLM, possibly quite soon. Claude Sonnet 4 and similar class models have enough reasoning performance, that agentic systems can be quite reliable. Which means we hit the base level of “reasoning” performance needed, and we can extend that “performance” in domain specific ways by lightly customising the agentic framework, with no need to fine tuning. The elimination of fine tuning to build domain specific agents is a huge game changer. But it also means that putting together a 10x or 100x efficient model, with “reasoning” performance equivalent to current gen LLM would also be a huge game changer. It opens up the possibility to apply this tech into spaces that currently require either lots of specialists knowledge to fine tune an LLM, or a huge amount of on tap compute to allow the agents to take enough turns to slowly “reason” they’re way through problems.
But a Claude Sonnet 4 that runs on a iPhone for example. That would make Apple’s complete failure to improve Siri look like a genius level move. Why bother with small incremental improvements using current tech, when waiting a few years, and just stuffing a full fat LLM and agent system into an iPhone will basically give you the ultimate Siri.
Right now they are trying to make sure the AI you run on your phone can't be used for coding or other tasks. They want to keep it expensive. They have no incentive to make it more efficient. No need.
The models aren't designed for efficiency either. There was a little scare with DeepSeek, but it turned out to be nothing. At the end of the day you need a lot of hardware and juice to run these things at scale. The costs aren't coming down any time soon. Nor is the impact to the environment unfortunately either.
The most logical future of AI is strictly on-device, capable enough to understand the user and making sense of the algorithmic-anarchy that is our phone.
THAT is something exciting and practical, that doesn't require magical-thinking to desire.
(Edited for tone)
Do they think they're secret sauce is UX? There's better editors out there now too.
You want to know what the hype train of Cursor was for? It was marketing for LLMs.
Now that all the providers have moved towards in-housing their coding solutions, the sweetheart deals are gone. And the wrapper goes back to "at cost" usage. Which, on paper should be less value / $ than any of the tiers offered by the providers themselves.
Whatever data they collected, and whatever investments they made in core tech remains to be seen. And it's a question of making use of that data. We can see that it is highly valuable for providers (as they all subsidise usage to an extent). Goog is offering lots of stuff for free, presumably to collect data.
One interesting distinction is on cursor vs. windsurf. Windsurf actually developed some internal models (either pre-trained or post-trained I don't know, but probably doesn't matter much) swe1 and swe1-lite that are actually pretty good on their own. I don't think cursor has done that, beyond their tab-next-prediction model. A clue, perhaps.
Anyway, it will be interesting to see how this all unfolds 2-5 years from now.
The difference is open-source doesn't exactly have a giant marketing budget. So most people haven't caught on yet, but they will.
Cursor is a solid tool but as best I can tell there’s not a ton there.
It's useful and has SOME value. Just not enough unfortunately.
This is history repeating itself. How could a company that raised so much money possibly compete with another that has the same ... arguably better ... product for less? Marketing budget? Hype? First to market? Sure. Absolutely. All those things do fade though.
We've seen this movie before. Remember Sublime Text? Remember what happened when Atom and VS Code came along? Fortunately Sublime Text didn't over raise (if they raised anything at all, I can't remember). Point is, people catch on and save money if they can. So they will do the same with Cursor.
Use open source editors that have the same features...better ones even. I'll argue that Roo Code is much much better than Cursor. I even like Windsurf better to be honest, but I wouldn't pay for either. I'll support open-source and save my money to pay the LLM.
Cursor is about to go the way of Sublime Text or Notepad++. Might keep some cult following, but it's market share will drop off a cliff.
It's ok. Their investors don't care! This was all to get people to use LLMs more. Their investors are fine with the sacrifice. That's all it ever was.
Early on Cursor added value by finding clever to integrate LLM into an IDE, which would allow single shot output of an LLM to produce something useful, and do so quickly. That required a fair bit of engineering to make happen reliably.
But LLM agents completely break that. The moment people realised that rather than trying to bend our tools to work within the limits of an LLM, we could instead just make LLM “self-prompt” their way to better outputs, Cursors model stopped being viable.
It’s another classic case of the AI “Bitter Lesson”[0] being learned. Throwing more data, and more compute at AI produces faster, better progress, than careful methodical engineering.
[0] http://www.incompleteideas.net/IncIdeas/BitterLesson.html
I've tried Aider and other agentic options and its amazingly clunky. Maybe I'm looking at the "apple vs linux" effect: I'm the apple user that just expect things to work out of the box, and although there are way better alternatives, the integration is worse.
This is also a bit of foreshadowing for all SaaS by the way.
Remember, nothing prevents anyone from simply vibe coding anything they see out there in the future. It turns the value of all software to near $0.
Cursor will go under. They'll be acquired for cheap (making more headlines, good press!), go bankrupt, or completely change their business into something else and undergo major restructuring. The point of Cursor is NOT to be a profitable ai code editor business. Look at the investors. Thrive dumped over $9B, using that $9B to advertise, market, and sell the product and services offered by the other companies they invest in -- the LLMs. At the end of the day, that's where the money flows.
In fact, none of the investors care one bit of Cursor survives. It's served its purpose. It was the "freebie" that people give out, the appetizer, for something bigger.
'We previously described ... as "rate limits", which wasn't an intuitive way to describe the limit. It is a usage credit pool'
Cursor just makes that window one month long.
Technically, that's a rate limit.
But yeah, only technically.
This all becomes very clear when you do something that feels like magic in Claude Code and then run /cost and see you’ve blown through $10 in a single hour long session. Which is honestly worth it for me.
Think about this one. They don't even tell you what your usage is! Look at Roo Code showing you the context usage and cost of each conversation. Features to compact the context. It's built around bringing awareness to the unit economics of AI and built to give users choice. The tools that work to keep users in the dark are serving someone else's interests.
Not because of Cursor‘s pricing, but because in the end Claude Code is unmatched.
For example they can react to in editor linter errors without running a lint command etc.
From the site : “Supermaven uses Babble, a proprietary language model specifically optimized for inline code completion. Our in-house models and serving infrastructure allow us to provide the fastest completions and the longest context window of any copilot.”
For $200/month you can get equivalent value to a team of engineers. Plan accordingly! The stack is no longer safe for employment. You need to move up to manager or move down to metal.
Why couldn't Claude do a managers job?
I see people mention converting old legacy code from an old language to something more modern. I've also seen people mention greenfield projects.
Anything other than this? I'm trying to bring this productivity to my work but so far haven't been able to replace a week of work in a few minutes yet
That was gemini-cli, I could see some mistakes on trial run so created a GEMINI.md with system prompt and project description (about 50 lines) which clarified some tricky source layout situations
Second run it was fine, ran for about an hour or so -- I had attempted to do it manually a while back but it started to look like it would take a week or two
EDIT: I haven't used Tailwind much but would something like this do what you're saying, or not really? https://www.loopple.com/tools/css-to-tailwind-converter
But this particular project is not like a standard site and the CSS is in small fragments across 100s files and uses constants for some things like color values in places too
In that Loopple example you can see the conversion uses the Tailwind arbitrary value notation, the -[], so background-color:#afa8af gets converted to bg-[#afa8af], but I wanted nearest pure tailwind class bg-zinc-400, the agent seems to work out color distance fine so does all that in one-shot too
I also saw that it was 0.8x the ‘credit cost’ thinking still that I had 500.
Now to learn that the 500 has gone and you get unlimited only on auto shows how easy it has been to misunderstand what they’re trying to say.
Also, I’ve no idea how to find out the cost of MAX. Especially as their web agent has the text MAX next to the selected non-max model.
Twitter VCs won't repent, once again.
$20 on API pricing is what Claude Pro will give you in a day. It doesn't matter how good cursor is, this is a massive limitation and price differential that they can't overcome. Even if they go with DeepSeek which is much cheaper, they are still significantly more expensive than a Claude subscription.
Also, that wonderful big beautiful bill in the US just limited what states and government can do to AI. It's a clear and open highway for Anthropic, OpenAI, Google, etc. They don't have to worry about any kinda rug pulling on Cursor or anything. They can literally compete with their "customers" and put their "customers" or "partners" out of business with zero consequences.
Cursor deserve the criticism, but its been pretty obvious since they introduced the new Ultra plan that we were going to see classic enshittification on the formerly premium options. Very frustrating for long term supporters especially.
Note that although this update apologises for the miscommunication (which looks more like deliberate obfuscation), the only option for getting what we paid for (I'm on the annual plan) is the menu setting above, which should be default on!