> We’ve grown to a team of over 300 engineers, researchers, designers, and operators
That last word, operators, I have seen used multiple times over the past couple of weeks to refer to managers and politicians. Is that the usage here too? If so, is this a new trend in the tech world? I’ve certainly heard of “political operators” in TV shows about Washington DC, but the usage in tech is new to me.
I've only heard it used in tech when you have actual operations, in my experience that meant lab managers and technicians. I'm not sure what it is supposed to mean in this context.
i think its a bit of a VC lingo that i dislike seeing adopted by actual... operators. it casts the world into two: either you own a business, or you operate one. that makes it look like a 50-50 choice that is usually a valid option for a privileged few. also if the world actually looked like that then we'd have a lot less building going on and there are too many VCs already.
instead I propose to call VC's "non-operator characters" and see how they feel about that
Prob burning through 200% of revenue which I've seen elsewhere. But they also probably spend a fair amount training their own model. I don't think it's foundation model. But it's pretty fair to assume that $1bn revenue is about $2bn to Anthropic/GPT/Grok
This article claimed they had single-digit monthly cash burn in August, when they had over $500M ARR (so let's say $41M monthly) and 150 employees. If that is true, they are spending way less than 200% of revenue.
"Anysphere runs pretty lean with around 150 employees and has a single digit monthly cash burn, a source tells me."
Doubt it. Especially when you realize the cost to the company for an employee is much more than just take-home salary. Healthcare, employer payroll taxes & such all add up. You could also argue wether deferred comp like stock options & RSUs are calculated as the cost. The employee's "comp package" often comes in at 2x or more of their base salary.
fwiw the current going rate for frontier agent labs and model labs is 50x. 30x is actually a discount presumably for size. obviously that can go down, but if you avoided investing based on multiples you were an absolute fool for the last 3 years.
I barely use the autocomplete features of Cursor, and for agentic coding, Claude Code blows Cursor Agent out of the water. I don't think Cursor has anything that cannot be replicated in a week or two other than the first mover advantage; certainly not an advantage that cannot be justified at 30B+ valuation.
they are building their own editor (granted they didn't do it from scratch); they do build their own models (see composer);
they may not have done a lot of this from scratch but there's still a lot of innovation in what they're doing. they're also building a pretty fantastic product and clearly the leader today in AI innovation for IDEs.
may not be everyone's cup of tea; but i think you might be detracting some of their innovation.
yeah its not fair to call it a finetune because finetune carries connotation of "there wasnt that much extra compute and data added". RLFT has a lot more added to it as Sasha alluded in his talk https://www.youtube.com/watch?v=md8D8eNj5JM - the x axis is log scale, think about that
the framing here is more about "why would you start from random weights when perfectly good starting weights exist" https://www.latent.space/p/fastai
It's more than that. They have both their own completion model and now agentic one. It's not a basic fine-tune, because it's faster than anything else available out there, so there's something interesting in the architecture itself.
Being this capital heavy also "justifies" their valuations. New shares issued in each funding round are typically around 10% of total shares, so to get to a valuation of $30B you have to raise something around $3B
Of course you could also just spend your money wisely and not do another funding round, but then how are people supposed to know how much you are worth? And how are investors supposed to know they made a great investment?
I am grateful to Cursor for ushering in the new age of coding beyond the "old" Github Copilot. I am also grateful to their VCs for subsidizing my coding. I am going to use their money to write subsidized code as long as the party lasts.
I like Claude Code in the terminal. For me it's so good it don't need IDE integration. I'm just using emacs and magit to navigate the code out of band.
I've spent quite some time evaluating the different tools over the last year, for both working on my employers complex codebase, and for my personal projects. At the start of the year I found Cursor pretty unsatisfactory and unable to complete tasks. At the time I rated Cline+vscode as the best agent and experience.
Now Claude code and Cursor are the best options imo, and I would say Cursor takes the edge for ide integration. Claude, as a separate thing to the ide, does mean you can do now flexible things like run it in a script loop.
Copilot doesn't get a shout in. It's fine for autocomplete but as a full agent it doesn't seem there yet.
If you're paying for it yourself Cursor seems to give the most bang for buck as well.
Comments like this remind me how much of the ai/agentic ecosystem is based on people's personal vibes and emotions
I've seen very little, meaningful difference. They all have their quirks, things their good at / bad at. The underlying models are very similar as well
I think the problem is that it can be a full time job on itself to try to test all of the available alternatives. The models and editors, cli tools that aims for "increasing developer productivity using LLMs" comes and goes much faster than most people can even track.
I think what you are saying is true too but another angle is that people use these tools in different ways so they yield different results. Hell even the expectations are different. Someone prompting for some React components will much happier with Claude sonnet 4.5 than me. I do heavy GPU programming and scientific computing stuff where LLM will mostly give you hallucinating answers 80% of time.
Okay, I'll delve deeper. You're talking about one thing, the models, which Cursor has a bunch available to choose from. Yes, I agree practically no difference there in terms of what junk is being spit out.
But code review is really important to me and nothing comes close to Cursor in regards to reviewing the code the LLM generates. I can keep the good parts, modify or throw out the bad parts. I can go back to a checkpoint easily when things get really bad. What solution comes even close to that? Cursor nails this really well. Claude code acolytes tell me to just use git commands. Yeah, no thanks.
you lost me here because this is based on your opinion and impressions, not data. How do you know nothing comes close?
The PR experience you describe is available in several options and setups. Different strokes for different folks, your choice is not superior by any meaningful measure other than your own preferences
I have both Cursor and VS code copilot in my work machine, but haven't really felt the need to use Cursor. VS code agent mode with Claude Sonnet is actually taking care of everything so far, plus I get to keep using my old launch config and debugging workflow.
> We’ve grown to a team of over 300 engineers, researchers, designers, and operators
That last word, operators, I have seen used multiple times over the past couple of weeks to refer to managers and politicians. Is that the usage here too? If so, is this a new trend in the tech world? I’ve certainly heard of “political operators” in TV shows about Washington DC, but the usage in tech is new to me.
I've only heard it used in tech when you have actual operations, in my experience that meant lab managers and technicians. I'm not sure what it is supposed to mean in this context.
Trickle-down titles? I’m familiar with operator to refer to tier-1 special operations personnel like Seal Team 6 and Delta Force.
I believe in their case it means operations people as in sysadmins. Ops like in DevOps.
HR, finance, sales any non engineer is called an operator these days basically
i think its a bit of a VC lingo that i dislike seeing adopted by actual... operators. it casts the world into two: either you own a business, or you operate one. that makes it look like a 50-50 choice that is usually a valid option for a privileged few. also if the world actually looked like that then we'd have a lot less building going on and there are too many VCs already.
instead I propose to call VC's "non-operator characters" and see how they feel about that
> Today, we’re pleased to announce a new round of financing: our Series D of $2.3B at a $29.3B post-money valuation.
> We’ve also crossed $1B in annualized revenue
A 30x revenue multiple on (presumably) relatively low-margin revenue is certainly punchy.
One wonders how much of their $1bn of ARR they're paying straight through to Claude/Anthropic.
300 employees, let's say average salary of 300k?
$90m in employee expenses so that's neglible.
Prob burning through 200% of revenue which I've seen elsewhere. But they also probably spend a fair amount training their own model. I don't think it's foundation model. But it's pretty fair to assume that $1bn revenue is about $2bn to Anthropic/GPT/Grok
This article claimed they had single-digit monthly cash burn in August, when they had over $500M ARR (so let's say $41M monthly) and 150 employees. If that is true, they are spending way less than 200% of revenue.
"Anysphere runs pretty lean with around 150 employees and has a single digit monthly cash burn, a source tells me."
https://www.newcomer.co/p/cursors-popularity-has-come-at-a
$300k most def too high
Doubt it. Especially when you realize the cost to the company for an employee is much more than just take-home salary. Healthcare, employer payroll taxes & such all add up. You could also argue wether deferred comp like stock options & RSUs are calculated as the cost. The employee's "comp package" often comes in at 2x or more of their base salary.
its higher
fwiw the current going rate for frontier agent labs and model labs is 50x. 30x is actually a discount presumably for size. obviously that can go down, but if you avoided investing based on multiples you were an absolute fool for the last 3 years.
They haven't built their own editor, they haven't built their own models; what have they actually built?
Well, they delivered something that is usable and useful for me and my team, and a lot of people I know, and I guess that’s what counts in business?
I barely use the autocomplete features of Cursor, and for agentic coding, Claude Code blows Cursor Agent out of the water. I don't think Cursor has anything that cannot be replicated in a week or two other than the first mover advantage; certainly not an advantage that cannot be justified at 30B+ valuation.
"not an advantage that cannot be justified at 30B+ valuation"
cannot -> can
(the extra negative negated your point)
That is all well and good, but I think it's a fair question in terms of valuation. What is their moat other than momentum?
They don't have one. People have been calling this out for a while.
They're also royally screwed since the IDE is going to cease being the place this work is done soon. Your VCS and org chat will be the new IDE.
I've turned off Cursor's autocomplete. Every interaction with it feels like two steps forward and two steps back.
Have you tried Zed? Cursor is terribly slow and buggy.
they are building their own editor (granted they didn't do it from scratch); they do build their own models (see composer);
they may not have done a lot of this from scratch but there's still a lot of innovation in what they're doing. they're also building a pretty fantastic product and clearly the leader today in AI innovation for IDEs.
may not be everyone's cup of tea; but i think you might be detracting some of their innovation.
I mean, they have built their own model: https://cursor.com/blog/composer
And presumably they'll use the funding to build more than just a modified VSCode.
most likely a finetune of existing model
As a user, I don't care.
Composer-1 is very good for routine code edits.
Claude and Gemini get pulled in for hard problems and architecture.
yeah its not fair to call it a finetune because finetune carries connotation of "there wasnt that much extra compute and data added". RLFT has a lot more added to it as Sasha alluded in his talk https://www.youtube.com/watch?v=md8D8eNj5JM - the x axis is log scale, think about that
the framing here is more about "why would you start from random weights when perfectly good starting weights exist" https://www.latent.space/p/fastai
It's more than that. They have both their own completion model and now agentic one. It's not a basic fine-tune, because it's faster than anything else available out there, so there's something interesting in the architecture itself.
Everyone starts with some building blocks, some much bigger than others in Cursor's case.
Something people want, apparently!
What they the money for?
The average series D is 50-100M. This is 2.3B.
I'm wondering if AI coding companies almost NEED to be this capital heavy to pay for the massive LLM costs.
Being this capital heavy also "justifies" their valuations. New shares issued in each funding round are typically around 10% of total shares, so to get to a valuation of $30B you have to raise something around $3B
Of course you could also just spend your money wisely and not do another funding round, but then how are people supposed to know how much you are worth? And how are investors supposed to know they made a great investment?
They propably burn something in the order of 50M-100M per month in LLM API costs for models like Sonnet 4.5. So the answer would be: Yes.
I am grateful to Cursor for ushering in the new age of coding beyond the "old" Github Copilot. I am also grateful to their VCs for subsidizing my coding. I am going to use their money to write subsidized code as long as the party lasts.
As joke goes: capitalism is the best communism. I wonder how many interesting small projects were created as a side effect of cursor funding.
I love Cursor.
I’m greedy to ask but is there a better alternative? Hard for me to imagine. I tried Copilot was no where near as good.
I like Claude Code in the terminal. For me it's so good it don't need IDE integration. I'm just using emacs and magit to navigate the code out of band.
I've spent quite some time evaluating the different tools over the last year, for both working on my employers complex codebase, and for my personal projects. At the start of the year I found Cursor pretty unsatisfactory and unable to complete tasks. At the time I rated Cline+vscode as the best agent and experience.
Now Claude code and Cursor are the best options imo, and I would say Cursor takes the edge for ide integration. Claude, as a separate thing to the ide, does mean you can do now flexible things like run it in a script loop.
Copilot doesn't get a shout in. It's fine for autocomplete but as a full agent it doesn't seem there yet.
If you're paying for it yourself Cursor seems to give the most bang for buck as well.
Comments like this remind me how much of the ai/agentic ecosystem is based on people's personal vibes and emotions
I've seen very little, meaningful difference. They all have their quirks, things their good at / bad at. The underlying models are very similar as well
I think the problem is that it can be a full time job on itself to try to test all of the available alternatives. The models and editors, cli tools that aims for "increasing developer productivity using LLMs" comes and goes much faster than most people can even track.
I think what you are saying is true too but another angle is that people use these tools in different ways so they yield different results. Hell even the expectations are different. Someone prompting for some React components will much happier with Claude sonnet 4.5 than me. I do heavy GPU programming and scientific computing stuff where LLM will mostly give you hallucinating answers 80% of time.
full agreement, the "my choice is the best" discourse has become quite tiring, I feel the same way about the rust stan'n
Okay, I'll delve deeper. You're talking about one thing, the models, which Cursor has a bunch available to choose from. Yes, I agree practically no difference there in terms of what junk is being spit out.
But code review is really important to me and nothing comes close to Cursor in regards to reviewing the code the LLM generates. I can keep the good parts, modify or throw out the bad parts. I can go back to a checkpoint easily when things get really bad. What solution comes even close to that? Cursor nails this really well. Claude code acolytes tell me to just use git commands. Yeah, no thanks.
Vibe coding? Yes, no meaningful difference.
> nothing comes close to X
you lost me here because this is based on your opinion and impressions, not data. How do you know nothing comes close?
The PR experience you describe is available in several options and setups. Different strokes for different folks, your choice is not superior by any meaningful measure other than your own preferences
I don't. I'm fishing for alternatives.
There are better ways to ask for alternative options and experiences
Cursor has the best tab-complete.
For agentic coding, people tend to prefer Codex or Claude Code, but I haven't heard many opinions about Cursor's new Composer yet.
I have both Cursor and VS code copilot in my work machine, but haven't really felt the need to use Cursor. VS code agent mode with Claude Sonnet is actually taking care of everything so far, plus I get to keep using my old launch config and debugging workflow.
Of course, Sourcegraph is better.
What does Amp do better than Claude Code or Codex? I find the concept pretty appealing but the pricing is a bit scary.
It has better search and context management under the hood, which matters for big companies.
But you can also see quantitatively that Sourcegraph produces the most accepted code: https://www.theinformation.com/articles/openai-catching-anth...
Well, that's certainly something I care about. I'll give it a try, thanks!
"Our in-house models now generate more code than almost any other LLMs in the world"
How do they know how much code is generated by other LLMs outside Cursor?
It's a vanity metric. And if it had any basis in reality, they would have mentioned it in the sentence.
can pretty much triangulate across openrouter x feedback from the top 3 model labs to compare with internal usage and figure that out
> Cursor should be a place where it’s impossible to write bugs.
Ha. Does anyone run a total on how much VC funding has gone towards this goal? In aggregate?
Crazy for a 2.5 year old company
Nothing to see here. Just your average 2.5 year old start-up worth nearly as much as Ford or VW.
"oday, we’re pleased to announce a new round of financing: our Series D of $2.3B at a $29.3B post-money valuation."
What does that mean? They got $2.3 billion of VC money and are now worth $30 billion?