What to do when a model lab buys your dev tool

You pay $20 a month for an editor [3] that lets you pick between four companies’ models [2], and that neutrality is most of why you chose it. Then one morning there is a blog post on the vendor’s own site saying the company has been acquired, and the buyer turns out to be in the business of building models itself [1]. Nothing on your screen is different. Your projects open, your subscription renews, the model dropdown still has the same names in it. The only thing that has changed is who benefits from which name you click.

This guide is about what to do in the weeks after that post, and it applies to any tool where the vendor sits between you and a model you pay for. The worked example is Cursor, because it is the clearest one available and because the documents are public, but the procedure is the same for an automation tool, a writing tool, or a support inbox that got bought by a company with its own model. If you are at a company with a procurement function, a security questionnaire and a lawyer who reads change-of-control clauses, this is too small for you; they will do a version of this with more paperwork. This is for people spending their own money on a seat or five.

Nothing breaks on the day of the announcement

On 14 August 2026, Cursor published a post saying “Cursor has officially been acquired by SpaceX,” completing a process that “started in April, when we announced our partnership with SpaceXAI to accelerate our model training efforts” [1]. The stated benefit to customers was economic rather than functional. Access to “the largest fleet of GPUs in the world” means, in the company’s words, “we can provide customers with more capable models at lower cost” [1].

Three weeks later, the model list is still plural. Cursor’s own documentation today lists Claude Fable 5.1, Claude Opus 5 and Claude Sonnet 5 from Anthropic, Gemini 3.1 Pro and Gemini 3.8 Flash from Google, three GPT-5.6 variants from OpenAI, and its own Grok and Composer models [2]. Nobody removed the competition. That is the normal shape of these deals, and it is why the announcement feels like a non-event.

It is also not the first time this category has changed hands. Cognition announced on 14 July 2025 that it had acquired the coding tool Windsurf, taking on its “IP, product, trademark and brand, and strong business” along with the team, at “$82M of ARR” and “350+ enterprise customers” [8]. Editors get bought. The tool you standardised on last year can have a different owner this year without changing its login screen, and the useful assumption is that yours will eventually be one of them.

Ownership shows up in the pricing page before it shows up in the product

The place to look is not the feature list. It is the billing documentation, because that is where a new owner expresses a preference without taking anything away from you.

Cursor now runs two separate usage pools. One is for its own models, described in the docs as “Significantly more included usage for Cursor Grok 4.6, Grok 4.5, and Composer 2.5” [3]. The other is for everything else, “The pool for third-party models, charged at the model’s API price” [3]. On top of that, “On Teams and Enterprise plans, third-party model requests include a Cursor Token Rate of $0.25 per million tokens” [3]. So Claude, Gemini and GPT are all still there, and on those two plans each request to one of them is metered at the provider’s own API price plus a quarter of a dollar per million tokens that requests to the house model do not carry.

Read that structure rather than the amount. The amount is small. The structure says the house model is the cheap default and every alternative is a line item, which means your budget makes the routing decision for you over the following year without anyone announcing a change. The plan tiers work the same way, at $20 a month for Pro, $60 for Pro Plus and $200 for Ultra [3], with the included usage on the first-party pool being what actually differs. Two further details are worth noting in passing: when you exceed included usage you “Continue at the same API rates with pay-as-you-go billing” or upgrade the plan [3], and choosing regional data residency adds “a 10% uplift on Model pricing for eligible Models” [3]. Meanwhile, “On Teams and Enterprise plans, Cursor Router picks the model for each Auto request based on your optimization mode” [2]. If you leave the automatic setting on, the owner picks.

Your data was part of the transaction

Every SaaS privacy policy you have ever clicked through contains a business transfer clause. An acquisition is the day it stops being boilerplate. Cursor’s says that “In the event of a merger, acquisition, restructuring, bankruptcy, or other corporate transaction, personal data may be disclosed to counterparties and advisers as part of due diligence or transferred as part of the transaction” [4]. Nothing about that is unusual or hidden. It is the standard term, and it means the account you opened with one company is now held by another, on terms the new owner can revise.

So re-read the two paragraphs that matter to you, on the day you hear the news, and save a dated copy. The first is training. Cursor’s current policy states “We do not use Inputs or Suggestions to train our models, or permit third parties to use them for training,” subject to three exceptions: content flagged for security review, content you explicitly report as feedback, and cases where you have explicitly agreed [4]. The second is retention, which is where most policies get vague. Cursor’s says the company “retains your personal data only for as long as necessary to operate the Service effectively and to support legitimate business needs such as legal compliance, safety, dispute resolution, and enforcement of our agreements” [4]. That is a purpose, not a number, and if a client contract obliges you to state a retention period, a purpose will not satisfy it.

The practical move is unglamorous. Save the policy as a PDF with today’s date in the filename, note the two clauses you rely on, and diary a re-read for your renewal date. You are not building a legal case. You are giving your future self a way to notice that something changed, because vendors are not obliged to tell you in a way you will see.

Portability is a question about files, not a feeling

The reason an editor feels sticky is rarely the editor. It is the accumulated configuration you would have to rebuild. The good news is that most of it is already text, and text moves.

Cursor’s project rules “live in .cursor/rules as .mdc files and are version-controlled” [5], which means they are in your repository and travel with it. Better still, the same documentation confirms that AGENTS.md works as an alternative: place it in the project root “as an alternative to .cursor/rules for straightforward use cases,” and nested copies in subdirectories “will be automatically applied when working with files in that directory or its children” [5]. AGENTS.md is an open format, described by its maintainers as “a README for agents,” used by more than 60,000 open-source projects and read by a long list of tools including OpenAI Codex, GitHub Copilot, Google Jules, Zed, VS Code, Aider, Devin and Cursor itself [6]. Writing your team’s standing instructions there rather than in a vendor-specific format costs you nothing today and removes a rebuild later.

Then be honest about what does not move. Your chat history does not. Your codebase index does not, and if you have spent months letting a tool learn a large repository, that is genuine switching cost you should count rather than wave away. Anything configured only in the app rather than in a file has to be recreated by hand. The test to run once, this month, is simple: ask what a new team member starting from a clean laptop and your repository would be missing. Everything on that list is your real lock-in, and every item you can move into a committed file is one you have taken back.

Keep a priced alternative, not a hypothetical one

“We could always switch” is not a plan until you have an account, a number, and one piece of real work run through it. Direct model access is the cheapest form of insurance here because it prices your fallback in public. Anthropic’s list prices today are $2 per million input tokens and $10 per million output for Claude Sonnet 5, $5 and $25 for Claude Opus 5, and $1 and $5 for Claude Haiku 4.5, with the Batch API taking 50% off both input and output tokens [7]. Those are the numbers you would be paying if the editor stopped being the sensible route, and knowing them turns a vague worry into a comparison you can actually make.

The other half of the plan is testing the house model rather than arguing about it. A new owner’s model getting cheaper inside the tool is not a trick played on you; it might genuinely be the right choice for your work, and the price advantage is real money. The only evidence that settles it is one representative task you were going to do anyway, run on the house model and on your fallback, judged on the output rather than on a benchmark either company published. Do that once a quarter and you will always have a current answer, which is more than most teams have when the question suddenly matters.

What still goes wrong

Every price and clause on this page is a snapshot verified on the date in the header, and the reason each one carries a citation is that the ones most likely to move are exactly the ones the acquisition put in play. Pools get renamed, surcharges get folded into plan prices, and a policy paragraph can be rewritten between your reading it and your renewal. Check the source before you make a decision worth money, including when the source is this guide.

Portability also solves a narrower problem than it appears to. Your files move, your code was always yours, and none of that helps if what you actually lose is a model that behaved the way your team had learned to expect. There is no export format for that, and the honest version of the switching plan includes a fortnight of being slower somewhere else. Weigh that against the surcharge before you decide the surcharge is intolerable.

Finally, this whole procedure has a bias, and the bias can cost you. Optimising for reversibility is right when you are spending your own money and cannot absorb a bad year, but a team that refuses to commit to anything never gets good at anything, and the deepest work usually happens inside one environment somebody has learned properly. An acquisition is a reason to check the exit, price the alternative and move your config into files. It is not by itself a reason to leave a tool that is doing the job.