The AI Editor That Thought It Was Neutral Just Learned It Isn't
For a few years now the story of AI coding tools has been that you no longer have to choose a single model. You fire up an editor, and inside it sit several of the frontier labs at once, Claude and GPT and Gemini and a couple of others, all reachable from the same box. The pitch was a kind of neutrality. One tool, every brain. It felt like the model market had matured to the point where no single company could lock you in, because the tool in front of you refused to be owned by any of them. That comfortable idea just took a hard blow, and the blow did not come from anything I would have expected, not a technical failure or a price war. It came from ownership. On November 12, OpenAI stopped supplying its models to Cursor, the most popular assistant in that category, and the reason it gave was that Cursor's parent company had been bought by a rival OpenAI does not want to do business with.
Cursor is worth slowing down on because it is the clearest example of the model neutral bet. It never trained a frontier model of its own. Its entire product was a harness around other people's models, and its differentiation was that it connected to all of them. For a developer that was the whole appeal. You were not choosing a model and its owner. You were choosing a tool that let you move between owners as the mood or the price shifted. The company behind it, Anysphere, was acquired by SpaceX in August for roughly sixty billion dollars. Two weeks later OpenAI invoked a change of control clause in its contract and told Cursor its models would be gone by November 12.
Here is the strangeness. The thing that was supposed to make the model market free and open, the neutral distributor, turned out to be the most exposed position of all. Cursor did not control the input it sold. It was a storefront renting shelf space from every lab, and the moment one of those landlords decided it disliked who owned the storefront, the shelf got emptied. OpenAI did not have to win a technical argument or a price war. It made a business decision about a counterparty, and an entire class of tooling had to rearrange itself around that decision. That is the uncomfortable fact at the center of this story. Access to the models you build on is not a technical guarantee. It is a commercial arrangement, and commercial arrangements can be terminated by someone who does not even use your product.
What is genuinely interesting, and what a lot of the coverage missed, is how little pain the cutoff actually caused. Cursor's cofounder said OpenAI's models were only about five percent of the editor's traffic. Anthropic, which had no reason to want the market to shrink, immediately said it would add compute to keep Claude flowing inside the tool. So the cutoff was a genuine disruption for the small number of users who were already wired to GPT, and barely a ripple for everyone else. That small number is worth sitting with, because it is the real market signal. A year or two ago, losing access to the most prominent lab's flagship would have been a catastrophe for any tool. Now it barely moves the needle, because the underlying models have become interchangeable enough that no single provider feels essential.
There is a clean parallel in business history, and it is not the one people usually reach for. Everybody points to the fights between app makers and app stores, and that is real, but the closer analogy is the retailer who built a national brand on shelves rented from manufacturers. As long as the manufacturers need the shelf, the retailer can negotiate. The moment a manufacturer decides the shelf is owned by a competitor, the retailer discovers it never really controlled its own inventory. Cursor is that retailer, and OpenAI is the manufacturer that decided its own competing storefront was the better place to sell. The lesson is not that neutrality is impossible. It is that neutrality is only as sturdy as the least motivated of your suppliers.
Before I lean on this too hard, I should grant the obvious objection. Cursor's situation is unusual because it involved an outright acquisition by a direct rival of a supplier. Most AI products will not have a SpaceX in their cap table, and most model providers will not have a reason to cut off a customer that is paying them. The change of control clause that OpenAI leaned on is standard in a lot of contracts, and it exists precisely for these moments. So it would be a mistake to overread one dramatic example into a general crisis. The model market is not about to shut its doors to resellers.
But set the drama aside and the underlying point survives, and it does not depend on the drama at all. Every AI product in the economy is built to some degree on models it does not own, and that dependence used to be treated as a technical detail, something you manage with API keys and rate limits. Cursor is the clearest proof that it is a strategic exposure. Whoever controls the model controls a slice of your product's existence, and they can exercise that control for reasons that have nothing to do with your product's quality. You can do everything right and still lose a feature, a customer, or a market because of a decision made in a boardroom you are not in.
And this is where the market seems to be heading, which is what I actually care about as a founder. If the models themselves are commoditizing, and this incident is decent evidence that they are, then the value in this industry keeps sliding toward the layers that are not so easy to cut off: the relationships with customers, the workflows and data that accumulate inside a tool, the trust that keeps someone coming back, and the distribution that lets a product reach people in the first place. Cursor losing five percent of its traffic to a supplier it is quarreling with is almost the whole story in miniature. The models were replaceable. The reason people stayed is that they were already inside, already wired into the way they work.
There is one more irony worth naming. Cursor built its reputation on being the impartial middleman, the one door that let everyone through. In the end that very impartiality is what left it exposed, because an honest middleman has no leverage. It owned the relationship with the developer, but it never owned the capability, and when a supplier decided to treat it as an enemy, the relationship did not help. The moat that looked broadest, being equally connected to everyone, turned out to be the thinnest, because it depended on the goodwill of everyone. You can see why the coding tools are racing to run their own models now, and why the base plans of the big editors point at their own or their parent's models first. Owning the supply is the only form of the moat no one can take from you.
I am not saying every AI business should run its own models or that dependence is fatal. Plenty of healthy companies will happily rent their models forever, and there is nothing wrong with that as long as the renter understands what it has signed. The mistake is to believe the lease is permanent, or that it is a technical matter rather than a business one. Cursor is a caution about a specific kind of confidence: the confidence that comes from being everybody's friend. It turns out that being everybody's friend is not the same as being indispensable to somebody. A tool can be loved and still be replaceable in the eyes of the companies that supply it. The models stopped being the scarce thing, and the only distribution you can call your own is the part a supplier cannot reach out and cut off with a contract.