How much did SpaceX pay for Cursor?

Approximately $60 billion, per public reporting on the June 2026 acquisition. This made Cursor (Anysphere) one of the most expensive AI-product acquisitions in history. The price reflects both Cursor's revenue (1M+ paying subscribers by early 2026 at $20-$200/month tiers) and the strategic value of its coding-data flywheel to SpaceXAI's model training.

Does Cursor still work with Claude and GPT after the acquisition?

Yes. Cursor continues to offer Claude Sonnet 5, Claude Opus 4.8, GPT-5.6 Sol / Terra / Luna, and Gemini 3.1 Pro alongside Composer 2 and Grok 4.5. There's no forced switch to Grok — the model picker is still open. However, Grok 4.5 receives preferential tuning (it was co-trained on Cursor data) and may get priority for new IDE-specific features.

What is co-training with Cursor data?

Co-training means Grok 4.5's post-training incorporated aggregated data from Cursor's users: accepted vs rejected AI suggestions, multi-file edit sequences, prompts that led to successful code changes, and debugging trajectories. This gives Grok 4.5 signal on 'what actually works in a real IDE' that models trained only on public code (GitHub) can't easily replicate. Data is aggregated and privacy-controlled per Cursor's ToS.

Quick Answer

What Is the Cursor Data Flywheel? SpaceX Acquisition Explained (July 2026)

Q: What is the Cursor data flywheel?

The Cursor data flywheel is the self-reinforcing loop where Cursor's ~2M developers generate real coding trajectories (which files they open, what edits they accept/reject, what prompts they use, which model outputs succeed), that data is used to train better coding models, better models attract more users, which generate more data. SpaceX's June 2026 acquisition of Cursor turned this loop into a proprietary training-data advantage — Grok 4.5 (launched July 8, 2026) was co-trained on aggregated Cursor traces.

Published: July 10, 2026

The Concept: Data Flywheel

A data flywheel in AI is a self-reinforcing loop:

More users → More usage data → Better model → More users
     ↑                                              ↓
     └──────────────────────────────────────────────┘

The most famous examples are:

Google Search: more searches → better ranking → more searches
Tesla FSD: more miles driven → more training data → better FSD → more customers
ChatGPT: more conversations → RLHF signal → better model → more conversations

Cursor’s version: developers using Cursor generate coding traces (which suggestions they accept, which they reject, what prompts they use, what multi-file edits succeed). That data is a uniquely valuable training signal for coding models — it captures “what works in production” that GitHub’s public code doesn’t.

Why the Flywheel Matters for Coding Models

Public code (GitHub) shows you what code exists. But it doesn’t tell you:

Which of these two implementations did a developer actually ship?
Which AI suggestion did they accept vs reject?
What prompt did they use to get the good version?
What did they type when the AI’s first attempt was wrong?
Which multi-file edit sequences produced correct working code?

Cursor’s telemetry captures all of this — anonymously and per its user agreement. That’s the training signal that generic GitHub scraping misses.

The June 2026 SpaceX Acquisition

June 2026: SpaceX announced the acquisition of Cursor (Anysphere) for a reported $60 billion. Context:

Cursor had 1M+ paying subscribers at Pro ($20/mo), Pro+ ($60/mo), Ultra ($200/mo)
Annual revenue was reportedly in the $400M+ range
2M+ total users
Growth trajectory was faster than GitHub Copilot’s early years

For SpaceX (already the parent of xAI as of February 2026), Cursor was:

A new revenue line (Cursor’s subscription business)
A distribution channel for Grok models (Cursor’s model picker)
Proprietary training data for coding models
A competitive moat against OpenAI/Anthropic on coding

The $60B price reflects all four, but the training data is the strategic prize.

The Result: Grok 4.5, July 8, 2026

Grok 4.5 was co-trained with Cursor coding traces. This showed up on launch:

Benchmark	Grok 4.5	Claude Opus 4.8
DeepSWE 1.0	62.0%	55.75%
Terminal-Bench 2.1	83.3%	78.9%
SWE Marathon (pass@1)	29.0%	26.0%

The gap over Opus 4.8 on these coding benchmarks is partly explained by the Cursor data advantage. Anthropic and OpenAI don’t have equivalent proprietary IDE telemetry.

Ecosystem Effects

For competing model providers (Anthropic, OpenAI, Google):

Cursor remains open to their models (Sonnet 5, Sol, Gemini)
But Grok 4.5 will get priority for new IDE-specific features
Cursor’s routing decisions become a competitive lever
The pressure to build their own IDE / data-flywheel grows

For competing IDEs (Windsurf, Zed, JetBrains Air):

Cursor now has a captive parent (SpaceX) with massive capital
Windsurf (Cognition) responded by doubling down on SWE-1.5 development
Zed continues to differentiate on performance (Rust, 0.4s startup)
JetBrains Air launched to consolidate their AI IDE story

For Cursor users:

No forced changes today — model picker still open
Grok 4.5 is the new default recommendation for coding
Composer 2 remains for precise diff work
Pricing structure unchanged (Free / Pro / Pro+ / Ultra)

The Regulatory Question

The consolidation has triggered DOJ antitrust review (per public filings). Concerns include:

Vertical integration: SpaceX now owns compute (Colossus), foundation model (Grok 4.5), IDE (Cursor), and distribution (X, grok.com)
Data-flywheel monopoly: Competitors can’t replicate the Cursor data advantage without building an equivalent IDE
Compute + model + distribution stack: Similar concentration as Google (Search + Chrome + Android) once faced

The review is ongoing as of July 2026. No immediate structural remedies have been announced.

What This Means for Developers

If you use Cursor:

Continue as normal — nothing has broken
Try Grok 4.5 as the new coding default (it’s the co-trained model)
Composer 2 still wins for precise multi-file edits
Watch for new IDE features that may be Grok-4.5-first

If you don’t use Cursor:

Windsurf, Zed, VS Code remain viable
Anthropic and OpenAI may accelerate IDE or agentic-tool investment
The pressure on other model providers to build data flywheels will increase

If you’re building an AI product:

Own your data flywheel — either build a native surface (IDE, app, tool) or partner with one
Public-data-only training is now table stakes; proprietary usage data is the moat
Vertical integration is a viable competitive strategy in 2026

Similar Plays to Watch

Other AI companies pursuing data flywheels:

Company	Flywheel Source
Anthropic	Claude Code CLI usage, Claude Projects
OpenAI	ChatGPT conversation data, Codex, Atlas browsing
Google	Search, Workspace, Chrome, Android usage
Microsoft	GitHub, VS Code, M365 usage
Meta	Muse Spark usage, Ray-Ban Meta glasses interactions
SpaceXAI	Cursor (coding), X (social), grok.com (chat)

Cursor’s acquisition was a signal that AI companies with existing large-model distribution now compete for proprietary training data as aggressively as they compete for compute.

The Bottom Line

The Cursor data flywheel is why Grok 4.5 beats Claude Opus 4.8 on three headline coding benchmarks despite being cheaper. SpaceX paid $60B for that flywheel — a signal about how much proprietary usage data is now worth in AI training.

For developers: Cursor is still open, still usable with any model. For the industry: proprietary data + vertical integration is now the competitive playbook.

Sources

SpaceX Acquires Cursor (SEC filings) — corporate announcement
Grok 4.5 Launch (SpaceXAI) — Cursor co-training mentioned
Cursor Company Blog — post-acquisition messaging
Anysphere Investor Announcement — acquisition details