Can AI Code Assistants be unlimited and afford vibe coding customers?

Calculating Cursor and friends’ runway. How much Vibe Coding can they survive?

Cursor has gotten a lot of heat these days, sticking to what seems illogical PR suicide.
This post tries to answer why they may have no choice but to make us angry, rug-pull and just quietly hope we’ll forget about it eventually.

Cursor changed its paid plans with the transparency of a lead box.
I think it sparked a wave, others followed and I think it’s only the beginning, in this post, I want to show you why.
As I mentioned on a reddit post a while ago, I think that’s because Cursor and co. pay per token, and those windows grow and also get more expensive with every new model.

I saw it coming somewhat, but wishful thinking made me avoid remembering why I went to Cursor in the first place.

I started with Windsurf. I liked it a lot more. Cursor lagged behind back then.
But I moved to Cursor because Windsurf changed their pricing to a request-based model, and at Cursor, I had unlimited requests. My experience wasn’t as good, but it was a way better deal.

And I wondered back then, and now too, how they can afford this? I’m here coding away, spending and, admittedly, wasting millions of tokens.

We see if, and how long they can afford this, and we’ll look at Bolt, Lovable, Replit and V0 too.

For this analysis I used the numbers officially provided by Lovable, Cursor, Vercel, Codeium/Windsurf and Replit.
The API Costs come directly from the model providers websites. I realize they likely have deals (and OpenAI now owns Windsurf) and pay less of course. We’ll make up for that by calculating distributions of usage scenarios.

I inferred MRR and ARR from one another depending on what was officially provided.
I.e. if Lovable announced ARR, I divided it by 12 to get MRR.
I made some assumptions.
I used 10k per month as salary per employee for all companies. I think that’s a fair average given that most employees are likely engineers. Though I understand that’s a high average salary.
I tried to be more generous to not make this a doomsday analysis. And I already don’t account for any other expenses.
I am making an educated guess that 60-80% are engineers and usually they make more than $ 100k a year, with 20-40% making less, maybe in part-time roles but also not a lot less than 100k per year.
The second assumption I made is that 30% of requests inside windsurf and cursor are for Claude sonnet 4. I also estimated 10% for Claude 3.7, with the remaining percentage split evenly among o3, o4, 4o, opus, gemini, and grok.
I found official information that Bolt, Lovable and V0 use Claude Sonnet 4.
I inferred the average token size per byte of code from several large datasets.

The last assumption I made is 100 employees for Cursor, the most recent official number I could find was around 65. But I heard a Cursor employee mention 100+ now, so I went with this number.
I’m happy to adjust it if someone can point out that it’s wrong.

Alright, to the juice.

I think very few people realize how expensive AI API Costs are and how they are charged.
API calls and requests are not per request but per token.

These are the prices I used from July 11, 2025 price in US$ per million tokens:

OpenAI - gpt-4.5: input 75$, output 150$
OpenAI - gpt-4.1: input 2, output 8
OpenAI - gpt-4o: input 2.5, output 10
OpenAI - gpt-o3: input 2, output 8
Anthropic - opus-4: input 15, output 75
Anthropic - sonnet-4: input 3, output 15
Anthropic - 3.7-sonnet: input 3, output 15
Google - 2.5-pro: input 1.25, output 10
Google - 2.5-flash: input 0.3, output 2.5
xAI - grok-4: input 3, output 15
xAI - grok-3-beta: input 3, output 15

In the graph below IDE represents Windsurf and Cursor, Vibe includes Replit, V0, Bolt and Lovable.

We try to cover the average user by Grid Search and to infer possible costs and revenue.

If a user makes 100 prompts per month, where each file has about 50 lines of code the API costs to Cursor and Windsurf would be about 1 Dollar.
With 2500 Lines of Code they cost Windsurf or Cursor 33 Dollars already, at 1500 prompts with an average size of 500 lines of code being read, it’s 200 Dollars.

It’s important to note here, that for Bolt, Lovable and others this cost is significantly less, these agents may touch more files and see more code but they also cost a lot less, so even minimal use of Opus (7% of the time) makes a huge difference.

Now that doesn’t seem like a lot, 200$ a month, but if you multiply this by the amount of paying users the providers have.

Well then these numbers get really large.

At a 1000 lines of code for an average project and 1500 prompts (around 50 per day, actually less, because tool calls), 360k users cost Cursor 72 Million $, PER MONTH.

This is a distribution and I think it’s a fair average, some might not use it at all, but we have had Vibe coders that pretty much are so vague the agent looks at 20 unnecessary files and creates 50 more, is extremely verbose and reads maybe 1000 of lines of debug output it created itself that tell the user nothing, the context eventually becomes so large, the user and the agent both forget what the point of most of the code it created in between was.

And it’s worth noting here, again, how badly Cursor and Windsurf scale compared to others, simply because of Opus.

Even if we’re very conservative, Cursor is not profitable at 20$ a month, even 200$.

And to drive this point home, look at how this scales without log scaling.

Now comes the fun part, let’s calculate how much runway each of the companies have or would have, depending on how people use it:

The runway calculations make it clear even Cursor, with over $1 billion in funding and $41.7 million in monthly revenue, can go bust if their average user is wasteful, or slightly productive. In the worst case, with the average user prompting 5000 times and 20’000 Lines of Code projects, Cursor’s runway drops to just 1 week.
That’s the dark red dots in the bottom left corner.

Notice how much steeper it is for Cursor than for example Bolt,

Or Replit:

Windsurf arguably has an even harder time:

This isn’t about a rug pull imho, the way they handled it, horrible.
But I am assuming they are paying or were paying API fee’s through the nose.
Especially after giving students free access in may.
Imagine all these Vibe Coders, Android and iOS Apps, the import statements are sometimes 50 lines of code already.

I think this is unsustainable, the Coding Agents are trying to win us over and winning the race by monopoly.
Maybe OpenAI (Google) and Windsurf can do this together.
But anyone that pays and doesn’t restrict, I don’t think it’s possible.

But even Bolt, V0, Replit and Lovable who restrict the users to cheaper Claude 4 (or 3.7) models can not be profitable if vibe coders go crazy.

And this is my opinion. You can only (and likely will) hit these high usage scenarios when vibe coding apps that you can’t build yourself.

Math is brutal. Lovable generates $5.8 million monthly but faces potential API bills of $32 million in high-usage scenarios. Windsurf earns $3.3 million monthly while potentially burning $2.96 billion on API costs alone.

The core paradox: Every feature that makes these products more useful (larger context windows, smarter suggestions, better code understanding) makes them more expensive to operate.
Tools and more documentation, rules, tests, automated tests.

I think that’s why, after Cursor, all of the others will slowly but surely increase their prices, unless, well since Windsurf and Openai/Google team up maybe Anthropic and Cursor can too?