Cursor Goes to SpaceX, Windsurf to Cognition: What Changes for Dev Teams

On June 16, SpaceX announced a $60 billion all-stock deal to acquire Anysphere, the company behind Cursor, with the transaction expected to close in Q3 2026.¹² SpaceX said the deal’s goal was to build “the world’s most useful AI models.”¹⁷ On the Windsurf side of this comparison: OpenAI entered exclusivity to acquire Windsurf in May 2025 for approximately $3 billion, before that deal collapsed when Microsoft’s partnership terms would have given it rights to IP from any company OpenAI acquired,³⁴ leaving the founding team to join Google DeepMind in a $2.4 billion deal in July 2025,⁵ while Cognition acquired the remaining company, brand, and roughly 210 employees at an undisclosed price three days later.⁶⁷ Cognition renamed the product Devin Desktop on June 2, 2026.⁸ Cascade, the agentic engine most developers associate with Windsurf’s core value, reaches end-of-life on July 1.⁸

Two of the three dominant non-Copilot AI code editors changed hands within months of each other. The questions that matter are not about deal valuations but about what each ownership change produced at the product layer, and what each means for teams currently standardized on either tool.

What Happened to Windsurf

The Windsurf story is not the “$250 million acquisition” it appeared to be from the headline. It is three transactions that landed within 72 hours of each other, plus a fourth that followed three weeks later.

OpenAI entered exclusivity to acquire Windsurf in May 2025 at an agreed price of approximately $3 billion.³ The exclusivity period expired in July 2025 when Microsoft’s partnership terms with OpenAI gave it rights to IP from any company OpenAI acquired. The carve-out attempt failed.⁴

Within hours of that announcement, Google DeepMind confirmed it was hiring CEO Varun Mohan, co-founder Douglas Chen, and roughly 40 senior engineers.⁵ The financial structure: approximately $1.2 billion in technology licensing fees from Google plus $1.2 billion in employment compensation packages for the hired team.⁹ Founders and the fund holding a roughly 20% stake were the primary beneficiaries; non-founding employees received minimal return from the deal structure given the liquidation preference waterfall.⁹

Three days later, on July 14, Cognition announced the acquisition of Windsurf’s remaining business: the product, IP, brand, and remaining employees, at a price neither company disclosed.⁶ Cognition’s own announcement described what it was buying: Windsurf had $82 million in ARR with enterprise ARR doubling quarter-over-quarter, 350-plus enterprise customers, and hundreds of thousands of daily active users.⁷

The post-acquisition timeline deteriorated quickly. Three weeks after the deal closed, Cognition laid off 30 of the acquired Windsurf employees and offered buyouts to the remaining 200-plus, with conditions attached to staying that included six-day work weeks and 80-plus-hour weeks.¹⁰ CEO Scott Wu put the company’s position in writing: Cognition does not believe in work-life balance.¹⁰

On the business side, the numbers held. Cognition raised $400 million at a $10.2 billion valuation two months after the acquisition, with combined ARR reported to exceed $155 million.¹¹ By April 2026, Bloomberg reported the company in talks to raise at a $25 billion valuation. The business is real, separate from the staff transition.

What Cognition built with the Windsurf product is a different question. On June 2, 2026, Windsurf was renamed Devin Desktop via an over-the-air update.⁸ The new interface ships an “Agent Command Center,” a Kanban-style multi-agent manager that replaces the original Cascade panel as the default surface. Cascade itself, the agentic code-execution layer that defined Windsurf’s product positioning, is replaced by Devin Local, a Rust rewrite that Cognition claims delivers 30% better token efficiency.⁸ Cascade dies July 1, 2026. Devin Desktop also added Agent Client Protocol (ACP) support, letting teams run other ACP-compatible agents (Claude Agent, Codex, OpenCode) from within the same IDE surface.⁸

Pricing changed in March 2026: free-tier model access was restricted, Pro moved from $15 to $20 per month, and a $200/month Max tier was added.¹² Developer community reaction to the free-tier restriction was sharp; the prior Codeium-era free-tier model access had been a specific reason individual developers chose the tool over Cursor and Copilot, and removing it without warning felt like a bait-and-switch.

The original engineers who built Windsurf’s architecture are at Google. The product has a new name, a new runtime, and a new interface. Teams choosing “Windsurf” today are choosing a different product from the one they signed up for, built by a different team under a different strategy.

What Is Happening to Cursor

The SpaceX deal is structurally cleaner than the Windsurf sequence: one buyer, founding team reported to stay post-close, single transaction.¹ The deal is not yet closed; Cursor continues operating under Anysphere management through Q3 2026.¹

The option structure matters here. SpaceX secured an option in April 2026 to either acquire Cursor for $60 billion in all-stock or pay a roughly $10 billion break-up fee.¹ That committed SpaceX to $10 billion in downside to maintain the right to buy at $60 billion. When the SpaceX IPO closed in June 2026, the option was exercised.¹

The rationale centers on compute: per SpaceX’s own announcement, it has spent recent months jointly training a model with Cursor on xAI’s Colossus infrastructure, set to ship in both Cursor and Grok Build.¹⁷ Cursor’s Composer 2.5 model, released May 18, 2026 and built on a Moonshot Kimi K2.5 checkpoint with a mixture-of-experts architecture trained with 25 times more synthetic data than Composer 2, is the last standalone Anysphere model before the Colossus-trained replacement.¹³ Secondary analyses of Cursor’s blog visual charts report Composer 2.5 at 79.8% on SWE-Bench Multilingual and 69.3% on Terminal-Bench; the official blog presents these comparisons as charts rather than text-parseable figures, so the specific numbers come from third-party writeups and not from primary text.¹³ No independent replication exists.

On model breadth today, Cursor supports six vendor families: Claude (Fable 5, Opus 4.8, Sonnet 4.6), GPT-5 through GPT-5.5, Gemini 2.5 Flash through Gemini 3.1 Pro, Grok 4.3 and 4.20, Kimi K2.5, and the Composer line (1 through 2.5).¹⁴ Cursor also shipped Cloud Agents in version 3.5 (May 2026), running in isolated cloud VMs with full desktop environments and browser access, capable of working across multiple repositories in parallel asynchronously.¹⁴ The three-mode architecture (Tab completion, targeted Composer edits, Agent Mode for directed tasks) that the in-house Composer model piece covered in May 2026 now has a fourth mode bolted on top.

The critical open question is model vendor continuity. Neither Cursor, Anysphere, SpaceX, nor xAI has issued any statement on whether Claude, GPT-5, or Gemini access continues after the acquisition closes.¹ The structural incentive post-close runs toward Grok and the Colossus-trained model. Whether existing enterprise data protection agreements tied to specific model providers survive unchanged through the ownership transition is not answered. Enterprise teams with contracts that specify Anthropic or OpenAI data handling terms should get written clarity on continuity before signing or renewing anything that spans Q3 2026.

JetBrains’ 2026 AI developer tool research found Cursor among the leading independent IDE options for developer satisfaction.¹⁵ No Windsurf or Devin Desktop figure appeared in the same research.

What the Tools Look Like Side by Side

Both are VS Code forks. Both support the Open VSX extension registry rather than the Microsoft-controlled Marketplace. The product philosophies have diverged, and the acquisitions accelerated that divergence.

No independent benchmark comparing the two exists. Cursor publishes no SWE-Bench score for any of its models. Cognition publishes results for its SWE-1 model family against a proprietary evaluation suite, not SWE-Bench Verified.¹⁶ There is no published head-to-head comparison on any shared benchmark. For what the SWE-Bench leaderboard actually measures and where its methodology breaks down, the explainer here is worth reading before using any vendor-reported score as a decision input.

One material Devin Desktop advantage: the Windsurf/Codeium extension layer has always supported VS Code, JetBrains, Vim/NeoVim, Xcode, and 40-plus other editors as plugins. Cursor is a standalone fork and cannot be loaded as a plugin into another IDE. For teams where the standard environment is JetBrains or where developers have established workflows in other editors, this distinction is not trivial.

For the three-way billing comparison from the April 2026 billing reshuffle and the earlier Windsurf Devin Review bundling piece that covered the pre-Devin-Desktop pricing breakdown, both document the landscape before the Cognition rebrand changed the SKU structure.

The Vendor Risk Calculus

Both acquisitions introduced distinct risk categories that teams should price explicitly at renewal time.

Model vendor continuity in Cursor. The tool currently supports six model families. Post-SpaceX close, the structural incentive runs toward Grok and the Colossus-trained model. No statement has been made about what happens to existing enterprise data protection terms tied to Anthropic or OpenAI. A team whose Cursor enterprise agreement specifies Claude data handling should verify whether those terms survive a SpaceX subsidiary before Q3 close.

Product continuity in Devin Desktop. Cascade ends July 1. Devin Local replaces it. Teams using Windsurf for Cascade’s specific flow architecture are migrating to a different runtime regardless of whether they stay on the subscription. Cognition’s 30% token efficiency claim for Devin Local is self-reported and not independently verified.

Procurement friction in Cursor. SpaceX’s association with Elon Musk and xAI creates procurement friction at organizations with vendor relationship policies. This is not a technical risk; it is a real-world enterprise sales obstacle that legal and procurement teams will raise, particularly in regulated industries. Developer community reaction to the acquisition explicitly flagged code data concerns under the new ownership structure.

Architectural knowledge loss in Devin Desktop. The engineers who designed Windsurf’s original codebase are at Google. The team that remained was offered buyout terms that most people would decline, and a substantial portion took the exit. Cognition has continued shipping updates (SWE-1, SWE-1.5, SWE-1.6, Devin Desktop), but architectural debt from founder departure typically takes months to surface in product quality. That window aligns roughly with now.

Which One to Choose

Choose Cursor if you prioritize current model flexibility across Claude, GPT-5, Gemini, Grok, and Composer simultaneously; if the three-mode UX with async Cloud Agents suits your team’s workflow; and if you can hold the model vendor continuity question open through Q3 2026. The founding team stays in place and the compute rationale is clear. The uncertainty is real but time-bounded. The full Cursor ARR and acquisition history covers how the tool reached this scale and what the SpaceX option structure signals about the deal’s logic.

Choose Devin Desktop if your use case calls for autonomous multi-agent task management over developer-in-loop editing; if the Kanban-style Agent Command Center fits your process better than Cursor’s sequential agent flow; or if you need JetBrains or Vim plugin support alongside AI features, which Cursor cannot provide. If you are currently using Windsurf, the Cascade-to-Devin-Local migration is mandatory regardless of any other decision, since Cascade stops working July 1.

Consider Claude Code if model lock-in is the primary concern. Claude Code’s terminal-native architecture and multi-agent patterns answer a different product question: autonomous task delegation over interactive editing, with the model vendor relationship explicit rather than embedded in an IDE abstraction. You know what you are buying.

Consider GitHub Copilot if enterprise procurement via Microsoft is the path of least resistance in your organization. The April 2026 billing reshuffle changed the cost math, but not the procurement story, which remains easier for organizations already inside GitHub Enterprise than any independent IDE offering.

For any of these choices, the benchmark that matters is not SWE-Bench. Run your own cross-file refactor, your own debugging session with real failing tests, and a multi-step task with ambiguous requirements from your actual codebase. The characteristic failure modes that distinguish these tools in practice, what the agent does with ambiguous requirements, how it handles naming convention drift across files, whether it introduces non-existent import paths on longer sessions, appear in 30 minutes of use on real code. They appear on no published leaderboard.

Frequently Asked Questions

Is Windsurf still available, or has it been replaced by Devin Desktop?

As of June 2, 2026, Windsurf was renamed Devin Desktop via an over-the-air update. The windsurf.com domain now redirects to devin.ai. Existing subscription billing and extensions continue under the new branding. Cascade, the agentic engine most developers associated with the Windsurf product, reaches end-of-life on July 1, 2026 and will stop functioning after that date. The replacement is Devin Local, a Rust-based rewrite developed by Cognition.

Will Cursor still support Claude, GPT-5, and Gemini after the SpaceX acquisition closes?

No official statement has been made on this by Cursor, Anysphere, SpaceX, or xAI as of June 20, 2026. The deal is pending Q3 2026 close. All major model families remain available today. The structural incentive post-close runs toward xAI’s Grok and the jointly-developed Colossus-trained model, but no removal has been announced. Enterprise teams with data protection terms tied to specific model providers should obtain written clarity on continuity before signing contracts that span the Q3 close.

What happened to the engineers who built Windsurf’s Cascade feature?

Windsurf’s co-founders, CEO Varun Mohan and co-founder Douglas Chen, along with approximately 40 senior engineers, were hired by Google DeepMind in July 2025 as part of a deal valued at approximately $2.4 billion total ($1.2 billion in licensing fees plus $1.2 billion in employment compensation). Cognition acquired the remaining Windsurf business and roughly 210 employees. Three weeks post-acquisition, Cognition offered buyouts to substantially all acquired staff under terms including 80-plus-hour weeks and six-day schedules. A significant portion accepted the exit.

How does Devin Desktop pricing compare to Cursor’s after the March 2026 changes?

Both charge $20/month at the Pro tier and $200/month at their respective power tiers (Cursor Ultra, Devin Desktop Max). Cursor also has a $60/month Pro+ tier with no direct Devin Desktop equivalent. The billing mechanics differ: Cursor uses a monthly credit pool that depletes on model calls, while Devin Desktop switched to daily/weekly quota refreshes in March 2026. Teams with bursty usage patterns hit these structures differently at the same headline price point, since Windsurf’s quota resets within the month while Cursor’s credit pool does not.

Is there an independent benchmark comparing Cursor and Devin Desktop?

No. Cursor publishes no SWE-Bench score for its Tab or Agent models. Cognition publishes results for the SWE-1 model family against its own proprietary evaluation suite, not against SWE-Bench Verified. No controlled third-party study comparing the two tools on identical tasks from the same repository exists. The comparison developers search for, a head-to-head on a shared external benchmark, has not been published.