Cursor SpaceX: What the Acquisition Changes in Your Editor (and What Nobody Confirmed)

The editor you opened this morning now belongs to the company that launches rockets.

Yesterday, August 14, 2026, Cursor published a thirteen-sentence announcement confirming it has been acquired by SpaceX. Thirteen sentences. They talk about GPUs, cheaper models, and the horizon. They don't talk about your code.

Here's the Cursor SpaceX dossier: what's confirmed in a primary source, what's press-only, what actually changes in the editor right now — spoiler: the most important change was already in the docs before the deal — and a checklist for anyone who depends on Cursor in production.

TL;DR

• What happened: Cursor was acquired by SpaceX on August 14, 2026, completing a process that started in April as a training partnership with SpaceXAI. Primary source.
• The price: the US$ 60 billion that made the rounds on your timeline isn't on any official Cursor page. That's press, not announcement.
• What changes in the editor: Grok 4.6 (August 12) is the flagship of the first-party pool; Anthropic's models sit under "Other Models" and are hidden by default in the official docs.
• The real risk: it's not the model menu. It's Cursor Router choosing for you in Auto mode.
• The hardest fact: Cursor has already admitted to contaminating a model's training with a snapshot of its own codebase — by accident.
• Deadline: the 50% discount on Grok 4.6 expires at the end of August 2026.

Cursor SpaceX: what was officially confirmed (and what only the press said)

Confirmed, primary source: "Cursor has officially been acquired by SpaceX." The acquisition completes the process that started in April with the SpaceXAI partnership. The technical promise is in there too, and the verb tense is worth reading: "We will have access to the largest fleet of GPUs in the world". Future tense. A promise, not delivered capacity. That's where the sentence that matters to your wallet comes from: "more capable models at lower cost".

Not confirmed, press only: the price. The all-stock US$ 60 billion shows up in Forbes, QZ, 9to5Mac, and another four outlets or so — but the announcement doesn't cite a single figure. ARR numbers, for their part, conflict wildly from source to source. If you're going to repeat this in a meeting, repeat it with the right label.

A detail on timing: on August 13, the day before closing, Cursor published three posts at once — the Firetiger acquisition, the AIUC-1 agent security certification, and cloud agents that are 3x faster. Stacked trust signals. Could be the calendar.

And one thing for the record before we go technical: anyone who tied their career to a specific editor found out yesterday that their career carries a risk item called M&A. Tools change owners; engineering depth doesn't — which is why at Clã Beer and Code the bar was never which IDE you open, but what you can build when the tool changes underneath you.

Native Grok 4.6: what changes about picking a model inside the editor

The announcement uses Grok 4.6 as its showcase: "Grok 4.6, which we released Wednesday". Wednesday was August 12, 2026 — two days before closing, not one, as some aggregators reported. The numbers, from consistent secondary sources: a 500K-token window, DeepSWE from 54 to 65.9, and APEX-Agents from 47.1 to 57.5 versus 4.5. A real jump on long agentic tasks. Not marketing.

Now the part most people didn't look at. Open the model documentation and look at the pools:

Cursor Models   →  Grok 4.6, Grok 4.5, Composer        (pool first-party, visível)
Other Models    →  Claude Sonnet 5, Claude Opus 5,
                   Claude Fable 5, Claude Sonnet 4.6   (hidden by default)

hidden by default. It's in the official docs. To use Claude in Cursor today, you enable it by hand, model by model. That was already true before the deal closed — the acquisition just removes the ambiguity about the incentive behind it.

Technical concept: Cursor doesn't sell models, it sells routing. Practical application: in Auto mode, the thing picking the model for each request is Cursor Router, an in-house classifier — "On Teams and Enterprise plans, Cursor Router picks the model for each Auto request based on your optimization mode". Auto is on by default on Teams. Product impact: Auto Cost mode has an obvious economic incentive to land on the first-party pool, where the margin is Cursor's, instead of burning a third party's API.

Translation: nobody needs to pull Claude from the menu to zero out Claude in practice. The default just has to stay on Auto. If you have a team on Teams, the Router mode is what you check today — not the menu.

"More capable models at lower cost": what that means for your plan

The announcement's promise has an expiration date. Literally.

Grok 4.6          US$  2 / M in   +  US$  6 / M out   (50% off, até fim de ago/2026)
Grok "fast"       US$  4 / M in   +  US$ 18 / M out
Claude Sonnet 5   US$  2 / M in   +  US$ 10 / M out
Claude Opus 5     US$  5 / M in   +  US$ 25 / M out
Claude Fable 5    US$ 10 / M in   +  US$ 50 / M out

Pegadinha: prompt >= 200K tokens no Grok 4.6 recobra o request INTEIRO
a US$ 4 / US$ 12 — não só o excedente.

All of these prices are in the model docs. Two practical takeaways:

1. The Grok 4.6 launch discount is 50% and it expires at the end of August 2026. The announcement's "lower cost" has two weeks to live. If you migrated a pipeline on price, redo the math in September.
2. The 200K gotcha is the one that hurts agents. Crossing 200K doesn't just make the overage more expensive — it makes the whole request more expensive, from the first token to the last. Anyone running agents with long histories feels it on the invoice before they feel it on the benchmark. If that's you, revisit your token-saving tactics before you grow the window.

The 3 unanswered questions: code privacy, team, and roadmap

1. Privacy. The thirteen sentences don't mention data, retention, training, or controller. Not once. And the silence carries weight because the track record is public: Cursor wrote, in the Grok 4.5 post, that the model was trained "jointly with SpaceXAI" using "trillions of tokens of Cursor data". The data flywheel doesn't start with the acquisition — the acquisition institutionalizes a flywheel that was already spinning.

That same post contains the most uncomfortable sentence in this dossier:

"Grok 4.5 has an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training."

Read it slowly. The company that controls the training pipeline leaked its own repository into the model without noticing, and only found out when the benchmark looked too good (the CursorBench scores were dropped from the launch). Credit where it's due: they published the mistake. But the engineering conclusion is inescapable — the guarantee that your repository doesn't take the same path is procedural, not technical. It depends on configuration and discipline, not on a barrier that prevents the accident.

2. Team. Zero public commitment on continuity, independence, or structure. The announcement is signed "Cursor Team", with no individual name.

3. Roadmap. Nothing here either. The signal is indirect: Firetiger brings agents that "monitor rollouts, catch regressions, investigate incidents" and feeds Cursor Origin, a git forge "for the agentic era". Cursor wants the whole loop — write, review, host, monitor — not just the editor. Bigger ambition, bigger lock-in surface.

Limitations and things to watch

Where you get burned:

• Privacy Mode is the line that separates everything, and it's your setting. Off, the data use policy authorizes using "codebase data, prompts, editor actions, code snippets" to train models. On, Cursor claims zero data retention with every provider (the page lists SpaceXAI, OpenAI, and Anthropic). If you've never opened that screen, you don't know which of the two worlds you're in.
• Grok Bot doesn't accept Legacy Privacy Mode. It's in the SpaceXAI docs: "Grok Bot requires data storage and does not support Legacy Privacy Mode". Anyone running in strict mode has to lower their privacy level to use it.
• Don't trust what circulated about "Colossus" and "3 ways they already work together" (Business Insider). I couldn't verify it in a primary source. Unconfirmed.
• There's a VentureBeat report that Anthropic cut xAI's access to Claude models after xAI engineers used Cursor as a route to consume Claude for competitive work. I couldn't open the source (429). If it holds up, the setup is bad on both sides: an editor with a structural incentive toward Grok, and a rival supplier with an incentive to make Claude harder to get.
• A date that went around wrong: aggregators say Grok 4.6 shipped on August 13. The official blog dates it August 12.

Should you stay on Cursor? A checklist for anyone who depends on it in production

This isn't the time for an emotional decision. It's time for an audit:

• [ ] Confirm your Privacy Mode state. If you don't remember turning it on, it's probably off.
• [ ] On Teams/Enterprise, check the Router mode. Auto is on by default. Choose between Auto Cost, Auto Balance, and Auto Intelligence — or get out of Auto and pin the model.
• [ ] Explicitly enable the models you want. Whatever is hidden by default won't show up on its own.
• [ ] Measure a month of real cost per model before you switch pipelines. Your September spreadsheet is a different one.
• [ ] Using Grok Bot? Document that the flow doesn't run in Legacy Privacy Mode. That's compliance, not preference.
• [ ] Have a tested plan B, not a hypothetical one. One day of work in a terminal agent or another editor already tells you whether the exit is viable — the 6-month comparison of Cursor, Claude Code, and Windsurf is the shortcut for that evaluation.

Outside regulated industries, running for the exit is overkill. In healthcare, finance, or government it's a different conversation: you have a new subprocessor in the chain and an announcement that said nothing about it. That becomes a contract review, not an opinion on Slack.

FAQ

Can SpaceX train a model on my code now? With Privacy Mode off, the policy already authorized using codebase data, prompts, and editor actions for training. On, Cursor claims ZDR with every provider. The acquisition doesn't change the policy text — it changes who's on the other side.

Is Claude going to disappear from Cursor? Nothing points to removal. But Anthropic's models are already under "Other Models" and hidden by default, and the Router in Auto has a cost incentive toward the house pool. Disappearing from the menu and disappearing from real usage are different things.

Is the US$ 60 billion official? No. It's press, with good consistency across outlets. The announcement cites no figure at all.

Is Cursor still worth it in 2026? Yes, once the audit is done. What changed isn't the quality of the product — it's the nature of the risk, which is now a governance risk. Migrating without a tested plan B trades a known risk for an unknown one. It's important to keep the two questions apart: here we're covering the risk created by the acquisition; the comparison of product, price, and output quality against competitors is in the 6-month comparison of Cursor, Claude Code, and Windsurf, which is the post dedicated to that choice.

What I'd take away from this week

The sentence that sticks isn't "SpaceX buys Cursor for US$ 60 billion". It's the one Cursor published in July, about a snapshot of the codebase that got into training by accident. The rest is corporate strategy; that one is data engineering in real life. And the blind spot was never the model menu: it's the Router, because a default is policy dressed up as convenience. What runs by default matters more than what's available.

I'll close on the irony of the timing, which showed up spelled out on Hacker News: "I use Codex for everything now". They bought the editor when a good chunk of the market had already moved to the terminal. If that read is right, the US$ 60 billion didn't buy the present of assisted development. It bought its rearview mirror.