If you don’t write software for a living, the name Cursor probably meant nothing to you until last week, when SpaceX agreed to pay $60 billion for it.
That is a strange sentence on its face. A rocket company bought a text editor, for the kind of money that buys a major airline outright, days after its own record IPO. SpaceX merged with Elon Musk’s AI company xAI earlier this year, and it framed the deal as a way to catch up to Anthropic and OpenAI in selling AI to businesses. Fair enough. But the more useful question is what you actually get for $60 billion here, because the answer turns out to be a clean little lesson in how this whole field works right now.
What is Cursor
Cursor is a code editor. More precisely, it is a fork of Visual Studio Code, the editor a large share of the world’s programmers already use, rebuilt so that an AI model sits in the core of it rather than hanging off the side as a plugin. You write code in it the way you always have, except you can also ask it questions about your project, hand it a feature to build, point it at a bug, and have it make the changes in plain language.
Two details matter for what follows. First, Cursor does not make the brain it runs on. You choose the model that powers it, Claude or GPT or Gemini or Cursor’s own in-house model, and you can switch whenever you like. Second, it reads your entire codebase and builds an index of it, so it understands how your files relate to one another rather than seeing only the one open in front of you.
It came out of nowhere. Anysphere, the company behind it, started in 2022, and by 2026 more than a million developers use Cursor, it has crossed roughly $2 billion in annual revenue, and it sits inside half of the Fortune 500. Among people who build software, it went from a thing you might try to the default in about eighteen months. That is the company SpaceX just bought.
Here is the question that should be nagging at you if you have ever used Claude or ChatGPT. You can already paste your code into a chat window and get good help with it. So what is the sixty-billion-dollar part?
What a harness is, and why a model needs one
The answer is a single concept worth carrying out of this post, because it explains a great deal more than Cursor. It is called the harness.
Start with what a raw AI model actually is, stripped of everything around it. It is passive. You give it some text, it gives you some text back, and then it forgets the entire exchange. It has no loop, no memory between calls, and no way to run code, open a file, or click anything. It is pure language and nothing else. Brilliant, and completely inert.
That is the thing you are talking to in a chat window. It can write you a flawless function and then do nothing further, because writing is the only move it has. You are the one who copies the code out, finds the right file, pastes it in, runs it, reads the error, and comes back to report what happened. The model supplies the thinking. You are its hands, its memory, and its eyes.
A harness is the software built around the model that supplies all of that for it. It is the scaffolding that turns a text generator into something that can actually do a job. A good one gives the model four things it does not have on its own:
- A loop. The harness calls the model, reads what it wants to do, carries it out (edits the file, runs the test, reads the error), feeds the result back, and decides whether to go around again. The model stops being a one-shot oracle and starts being a worker.
- The right context. Rather than dumping your whole project on the model, which actually makes its answers worse, the harness feeds it the relevant slices, found through that codebase index. Choosing what the model should look at is a real skill in its own right.
- Memory. Task lists, notes, and saved state, so a long job survives the model forgetting everything between calls.
- Tools. A terminal, the file system, web search, connections to your other software, so the model can take real actions instead of only describing them.
Cursor is a coding harness wrapped around a frontier model. If the model is the engine, the harness is the rest of the car: the wheels, the steering, the dashboard, the thing that turns raw power into somewhere you can actually go. We wrote about the individual parts of that car, the loop and the tools and the memory and the guardrails, in The Agent Vocabulary.
The harness often matters more than the model
This is the part the field landed on over the last year, and it is the part worth holding onto.
Anthropic, the company that makes Claude, ran an experiment on exactly this. It took Opus 4.5, one of the strongest models available, put it in a loop across many context windows, and gave it a high-level instruction: build a working clone of a real web app. On its own, even running continuously, the model could not do it. It would try to build everything at once and run out of room, or a later run would glance at the half-finished project, decide the work looked done, and quit. The model was plenty capable. What it lacked was structure around it. Once it was wrapped in a real harness, one agent to lay out the work, another to build it feature by feature, progress written to files, testing required before anything counted as finished, the same model started shipping the app.
The lesson people keep drawing from this is blunt: a well-built harness around a slightly weaker model will usually beat a poorly-built harness around a stronger one. The harness decides what the model can see, what it is allowed to do, and what happens when something breaks. Most agent failures are not the model reasoning badly. They are the harness handing it the wrong context or letting it run in circles.
Which reframes the whole business. The intelligence is rented. Any of these tools can call Claude or GPT, and the labs will sell that intelligence to anyone with a credit card. The hard, durable, valuable work is the harness built around it.
Why this is not the same as Claude Code
It is tempting to lump every AI coding tool into one bucket, but the distinction worth making is the surface each one gives you, because they are built for different jobs and do not really compete for the same slot.
A chat window, the Claude or ChatGPT tab you already use, is for thinking. Planning, architecture, learning, getting unstuck on an idea. It is not wired into your codebase, and it is not where you build a real multi-file project. Keep using it for what it is good at.
Cursor is a build surface. It is a full visual editor: a file tree, a terminal, your code in front of you, and every change the AI proposes shown inline for you to approve or reject. The amount of freedom you give it is a dial you control, from small suggested edits up to letting an agent run a whole task. That visibility is the entire point. You are never left guessing what the thing did to your code.
Claude Code, Anthropic’s own tool, is a leaner agent that lives in the terminal. Less to look at, more autonomy, built for handing over a big chunk of work and trusting it to run. It is where people tend to graduate once they know what good output looks like and want to delegate more of it. Both Cursor and Claude Code are harnesses around a frontier model. They are simply shaped differently, one visual and supervised, one lean and autonomous, for different moments in the work.
What it tells us
Step back, and the $60 billion stops looking absurd.
Notice what SpaceX did not buy. It did not buy a model. Cursor does not make the intelligence it runs on, and it lets you swap between the labs at will. The model is a commodity you rent by the token. What is worth $60 billion is the harness, the years of working out exactly what a coding model needs to see and how to put it there fast, along with the million-plus developers already living inside the tool every day and the steady record of how they work.
If that is right, it points somewhere specific. The chat box, the text field we all type into, starts to look like a transitional form, the way the dial-up modem was. It is how most of us first met these models, but it is a narrow slot, and the model wants to be everywhere the work already happens. Cursor is what it looks like when the model climbs out of a separate tab and into the thing you are actually doing. There is no reason that stops at code. The same move is already arriving in the document, the spreadsheet, and the inbox. We wrote about where the agents that depend on these harnesses actually stand today in The Honest State of AI Agents.
A rocket company paying $60 billion for a text editor reads as a punchline right up until you notice it bought the better car built around an engine it can rent cheaply. Then it looks less like a joke and more like one of the clearer bets anyone has placed on where all of this is going.