SpaceX Owns Cursor: What Changes for Developers

SpaceX now owns Cursor. The all-stock merger with Anysphere, Cursor's parent company, took effect on August 14, 2026, at an implied equity value of $60 billion, according to SpaceX's SEC filing. For developers, the editor still works the same way it did last week. What has changed is who owns it: one of the best-known independent AI code editors now belongs to a company that trains its own frontier models, and those models, Grok 4.5 and Grok 4.6, are already built into how Cursor picks a model for you. You can still use Claude, GPT and Gemini in Cursor. Keeping that choice open is now up to you.

This article covers what the deal actually is, what has visibly shifted inside Cursor since April, and a practical checklist for keeping your team's workflow model-agnostic, so that a future pricing or routing change inside Cursor is an inconvenience, not a migration.

What exactly did SpaceX buy, and when did it close?

The deal was built in three steps, each on the public record.

• April 21, 2026: a training partnership with an option attached. Cursor announced it was partnering with SpaceX on model training, using SpaceXAI's Colossus infrastructure because, in Cursor's words, it had been "bottlenecked by compute." CBS News and Yahoo Finance both reported that the agreement gave SpaceX the right to acquire Cursor for $60 billion later in the year, or to pay $10 billion for the collaboration instead.
• June 16, 2026: the option is exercised. SpaceX, a merger subsidiary called X67 Inc., and Anysphere signed an Agreement and Plan of Merger. The 8-K filed that day describes an all-stock deal at a $60.0 billion implied equity value, expected to close in the third quarter of 2026 once regulators approved it. CBS News quoted Cursor CEO Michael Truell: "SpaceX has exercised their option to acquire Cursor in an all-stock transaction with the goal of building the world's most useful AI models."
• August 14, 2026: the merger takes effect. According to the closing 8-K, Cursor's common and preferred shares converted into the right to receive 389,289,254 SpaceX Class A shares, and vested Cursor RSUs (restricted stock units, a form of employee equity) converted into another 1,752,426 shares. SpaceX also took on roughly 29.1 million unvested RSUs and about 44.4 million stock options. Cursor wrote on its blog the same day: "Cursor has officially been acquired by SpaceX."

Some context on the buyer. xAI, the company behind Grok, was folded into SpaceX earlier in 2026, according to TechCrunch and Yahoo Finance's reporting on the February announcement, and its models now ship under the SpaceXAI name. So Cursor's new parent is not just a company that rents out GPUs. It is a model maker that competes directly with Anthropic, OpenAI and Google, whose models make up much of Cursor's model picker.

Is Grok now the default model in Cursor?

Not quite, but it is now built into the default path. Three changes stand out.

Grok models are listed as Cursor's own. Cursor's models documentation groups Grok 4.6, Grok 4.5 and their Fast variants under Cursor's native models, described as jointly trained by Cursor and SpaceXAI, alongside Cursor's in-house Composer 2.5. Grok 4.5 launched in July 2026; Grok 4.6 followed on August 12, two days before the deal closed. Both are priced at $2 per million input tokens and $6 per million output tokens as of September 2026, according to Cursor's launch posts. Cursor says Grok 4.6 matches OpenAI's GPT-5.6 Sol on the Artificial Analysis Intelligence Index. That is Cursor's own claim, not an independent result.

The router depends on Grok. If you pick "Auto," Cursor Router chooses a model for each request, in one of three modes: Intelligence, Balance or Cost. Cursor's available-models help page calls Grok 4.5 "Cursor's flagship model" and lists it as required: "The router needs a powerful, cost-efficient model in the pool; blocking Grok 4.5 disables the router." Put plainly, an enterprise admin who blocks the parent company's model also loses Auto mode.

The routing logic is Cursor's to tune. An August 6 Cursor post on how the router works describes a complexity predictor called Compass, followed by a task classifier. It says Grok "offers strong value across broad, routine work," while GPT-5.6 Sol handles planning and codebase comprehension, Claude Opus handles execution-heavy work, and Claude Fable handles debugging and visual implementation. As published, that is a performance-based split, and nothing in it shows Grok being favored for business reasons. But the weights are proprietary, and the company that sets them now earns money when Grok is chosen.

For balance: Cursor's model list, as of September 2026, still includes models from Anthropic, OpenAI and Google, plus Z.ai, Moonshot and Meta. You can still choose any of them by hand from the model selector, or cycle through them with Cmd + /. Nothing has been taken away. What has changed is the incentive behind the default.

Why does ownership matter if nothing has been removed?

An independent tool sells access to other companies' models. A tool owned by a model maker is also a way to distribute that maker's own models. Those are different businesses, even when the product looks identical on day one.

The pressure points are predictable, even if none of them has happened yet:

• Pricing. Subscription usage that stretches further on in-house models than on third-party ones is ordinary economics for an owner. Cursor already gave launch-week bonus usage on Grok 4.5 and 4.6. That's a standard promotion, but it shows the lever exists.
• Feature timing. New agent features tend to be tuned first on whichever model the vendor knows best. Expect Grok to get the most attention.
• Supplier relationships. Anthropic, OpenAI and Google now supply models to a product owned by a direct competitor. None of them has announced a change in terms, but that relationship is now on different footing.
• Data governance. Cursor's data-use page (last updated September 3, 2026) says Privacy Mode prevents training on your code and that Cursor holds zero-data-retention agreements with all model providers. The page now links to SpaceXAI's own zero-retention FAQ. The protections are the same. What's new is that one of the model providers is now Cursor's parent company.

The GitHub Copilot comparison is useful. Copilot has always belonged to a company, Microsoft, with a close model partner, and teams took that into account when they chose it. Cursor's appeal, compared in our breakdown of the AI coding tool pricing wars, was partly that it had no such tie. That argument no longer holds, and procurement teams should re-score it rather than assume it still does.

For balance, the case for the deal is real. Cursor said openly in April that compute was its bottleneck. It now has, in its own words, "access to the largest fleet of GPUs in the world." A coding model trained with that much compute and priced at $2 per million input tokens is good for developers. Yahoo Finance reported Cursor's annualized revenue at about $2.6 billion when the deal was announced, so this is a business with paying customers to keep, not a demo to shut down.

How locked in are you today?

Not every part of a Cursor workflow is equally portable. Here is how each piece stands as of September 2026, based on Cursor's documentation:

Workflow layer	Where it lives	Portable to other tools?	Action
Project instructions	AGENTS.md or .cursor/rules/*.mdc	AGENTS.md: yes. .mdc rules: Cursor-specific	Make AGENTS.md your main instructions file
Tool and data connections	MCP servers	Yes, as an open standard	Build integrations as MCP servers, not editor plugins
Chat and agent model	Model picker, Auto or manual	Choice is open; Auto's routing is not	Pin models for tasks where consistency matters
Your own API keys	Settings > Models	OpenAI, Anthropic, Google, Azure OpenAI, AWS Bedrock	Keep a direct provider contract as a fallback
Tab autocomplete	Cursor's built-in models only	No	Accept it as the one fully Cursor-specific layer
Cloud agents and history	Cursor's infrastructure	Limited	Keep specs and decisions in the repo, not in chat

How do you keep a Cursor workflow model-agnostic?

None of this means leaving Cursor. It means making sure leaving would be cheap if you ever needed to. Five steps, in rough order of payoff:

1. Move instructions into AGENTS.md

AGENTS.md is an open markdown format for guiding coding agents, now stewarded by the Agentic AI Foundation under the Linux Foundation. It is supported by Cursor, OpenAI Codex, Google's Jules and Gemini CLI, GitHub Copilot's coding agent, Zed, Warp, Windsurf and others. Cursor's rules documentation confirms it reads AGENTS.md in the project root and in subdirectories, and that rules are added to the model's context in Agent chat (though not in Inline Edit or Tab). Keep build commands, test commands, conventions and architecture notes there. Save .mdc files for things only Cursor needs, such as rules that apply only to certain file paths.

2. Build integrations as MCP servers

The Model Context Protocol is an open-source standard for connecting AI applications to tools and data. Cursor, VS Code, Claude and ChatGPT all support it. An internal MCP server for your ticket tracker, database or deploy pipeline works in any of them. We covered why this matters in the AI agent protocol landscape.

3. Pin models on purpose

Auto is fine for exploratory work. For tasks where you need consistent output, such as code review passes, migrations or generated tests, choose a specific model by hand and write the choice into AGENTS.md or your team's runbook. That way a routing change can't silently change behavior you depend on. Team admins can also switch router visibility to "Displayed" in the admin dashboard so members can see which model handled each request.

4. Keep a direct contract with at least one provider

Cursor supports bring-your-own API keys (BYOK) for OpenAI, Anthropic, Google, Azure OpenAI and AWS Bedrock. Two caveats from Cursor's docs: your own keys only work for chat models, not Tab completion, and Cursor's zero-retention agreements don't apply when you use them. Your provider's own policy governs the data instead. A direct contract also means a second tool, such as a terminal agent, can take over immediately. Our Claude Code review covers one of those options.

5. Keep a second, open fallback ready

Know which open-weight model you would run if every hosted option got worse at once. Our mid-2026 open-source LLM survey covers the options. You probably won't need it, but thirty minutes of testing now is cheaper than a rushed evaluation later.

What should you watch next?

Four signals will show whether ownership changes how the product behaves:

• Router behavior over time. If Auto starts sending most complex work to Grok without published benchmarks to support it, that suggests business priorities are shaping the routing.
• Plan pricing. Watch for differences in how far your plan's usage goes on in-house models compared with Claude and GPT.
• Third-party model timing. Watch whether new Anthropic, OpenAI and Google models keep reaching Cursor as fast as they do other editors.
• Data-use page changes. The September 3 update already mentions SpaceXAI. Read each future revision.

The bottom line: Cursor is still a strong editor, and it now has more compute behind it. But it is no longer a neutral broker between model makers, and a tool's incentives matter more than its launch-day promises. Teams that keep instructions in AGENTS.md, integrations in MCP and a direct provider contract on file can stay with Cursor on its merits, and leave cheaply if that ever changes.

Frequently Asked Questions

How much did SpaceX pay for Cursor?

SpaceX's SEC filings put Cursor's implied equity value at $60.0 billion, paid entirely in SpaceX Class A stock rather than cash. Cursor shareholders received the right to 389,289,254 SpaceX shares, with more shares, RSUs and options for employee equity. The merger was signed on June 16, 2026 and took effect on August 14, 2026. Cursor now operates as a wholly owned SpaceX subsidiary.

Can I still use Claude, GPT and Gemini models in Cursor?

Yes. As of September 2026, Cursor's model list includes models from Anthropic, OpenAI and Google, plus Z.ai, Moonshot and Meta, and you can select any of them by hand from the model picker. The Grok models are listed as Cursor's native models, and the Auto router requires Grok 4.5 to be available.

Does SpaceX get to train on my code?

Cursor's data-use page, updated September 3, 2026, says that with Privacy Mode on, your code is not used for training, and that Cursor holds zero-data-retention agreements with all model providers, including SpaceXAI. With Privacy Mode off, Cursor may use code data to improve its features and train models. Check which mode your team's accounts use.

What happens if my company blocks Grok in Cursor?

On enterprise plans, admins can block specific models. According to Cursor's help documentation, blocking Grok 4.5 disables Cursor Router, so Auto mode stops working for that team. Other models stay available for manual selection, so developers can keep working by choosing models themselves.

Should my team switch away from Cursor?

Not because of the acquisition alone. The product hasn't removed any models or features. The practical response is to reduce switching costs: keep project instructions in AGENTS.md, build integrations as MCP servers, pin models for critical tasks, and keep a direct API contract with at least one provider. Then decide based on how the product performs over the next few quarters.