TL;DR
- SpaceX (via SpaceXAI, formerly xAI) acquired Cursor — the AI-first code editor with $4 billion in annual revenue — for $60 billion in all-stock. The deal closed August 14, 2026. Cursor is now a wholly-owned subsidiary of SpaceXAI. Architecture: Cursor is a VS Code fork built by Anysphere, founded 2022 by four MIT students. - On August 28, two weeks after closing, OpenAI notified SpaceX that it would terminate Cursor's access to its models on November 12, 2026 — next week. OpenAI models currently account for approximately 5% of Cursor usage. Cursor still offers models from Anthropic, Google, and SpaceXAI (Grok). SpaceXAI and Cursor are preparing their first jointly developed model for imminent release. - Meta acquired Conductor, an AI coding-agent startup, in a separate deal. Conductor built autonomous code-deployment agents rather than a code editor — a different product layer from Cursor. - Thinking Machines Lab released Inkling 2 Small on November 5: a 594B-total / 17B-active-parameter MoE with native vision and variable reasoning effort, trained on 12 trillion tokens in 68 hours using RL on long-horizon coding tasks. Artificial Analysis scores it 31 on its Intelligence Index — the highest-scoring open-weight model developed outside China if weights ship next week under Apache 2.0. SWE-Bench Pro v2 (Hard): 86.3%.
The AI coding tools market has been moving fast enough that the week's headline — "xAI acquires Cursor for $60B" — requires clarification: the acquisition closed in August. The news this week is that OpenAI is ending Cursor's access to its models on November 12.
That is the actionable item for teams using Cursor with OpenAI models. Next week, the GPT-4o, GPT-6 Sol, and GPT-6 Luna options in Cursor's model selector will stop working.
The Cursor/SpaceX acquisition and what it means
Cursor was the most popular AI-first code editor in 2026. It is a fork of Visual Studio Code — the most widely used code editor in the world — with AI assistance integrated at every level: autocomplete, inline chat, code generation, codebase-wide context understanding, and agent-mode task execution. Its $4 billion in annualized revenue at acquisition reflected genuine enterprise adoption, not speculative valuation.
SpaceX secured the option to acquire Cursor in April 2026, with an alternative: pay $10 billion to work together without full acquisition. It exercised the $60 billion purchase option in June. The deal closed August 14. Cursor became part of SpaceXAI — the entity formed when SpaceX absorbed xAI (which had already acquired X in 2025) in February 2026.
The strategic logic is clear: SpaceX is building Colossus, a massive AI computing facility in Memphis, Tennessee. Cursor provides access to what SpaceX described as a "wide distribution to expert software engineers" — the customer base that any frontier AI company needs to demonstrate utility to. Getting Cursor's users onto Grok models, and using their coding sessions to improve Grok's coding capability, creates a data flywheel that neither OpenAI nor Anthropic can replicate from outside.
OpenAI's response to the acquisition was deliberate: it waited until two weeks after closing to announce that Cursor would lose access to OpenAI models on November 12. The timing suggests OpenAI's terms of service or commercial agreements with Cursor prohibited model access from competitors' subsidiaries, and that OpenAI chose to enforce this after the deal was legally complete rather than before.
Cursor CEO Michael Truell has stated that OpenAI models account for approximately 5% of Cursor usage. This is a low enough number that the technical disruption is limited — the other 95% of model usage (Anthropic Claude, Google Gemini, Grok) is unaffected. But the optics matter for enterprise customers who made procurement decisions assuming Cursor provided neutral access to multiple model providers. It is no longer neutral: it is a SpaceXAI product that happens to resell Anthropic and Google models.
SpaceXAI and Cursor are preparing a jointly developed model using the Colossus supercomputer infrastructure. The model's release has been pushed back multiple times for efficiency optimization. When it ships, Cursor becomes the primary distribution channel for SpaceXAI's first native coding model.
Meta's Conductor acquisition
Meta acquired Conductor separately — a coding-agent startup focused on autonomous code deployment rather than a code editor. Where Cursor operates at the developer's side (helping a human write code), Conductor's agents operate further down the pipeline (autonomously deploying, testing, and iterating on code with minimal human direction).
The two acquisitions cover different layers of the same workflow: SpaceX/Cursor owns the human-in-the-loop coding interface; Meta/Conductor owns the downstream autonomous execution layer. Both are bets on the same end state — most code gets written and deployed with minimal human authorship — but from different entry points in the development pipeline.
Together they represent a consolidation of AI coding infrastructure into the same companies competing at the frontier model layer: SpaceX (xAI), Meta (Muse), Google (Gemini), Anthropic (Claude), and Microsoft (GitHub Copilot, plus the MAI-Base-2 internal track). The AI coding tools market is becoming a distribution channel for each lab's models, not a neutral marketplace of tools.
The enterprise action item for Cursor users
If your organization uses Cursor and has workflows built on OpenAI models accessed through the Cursor interface, those workflows will fail on November 12. The fix is straightforward: switch the model selection in Cursor to Claude, Gemini, or Grok before November 12.
The more significant governance question: did your procurement team evaluate Cursor as a neutral developer tool or as a product from an AI lab? If the former, the August acquisition changed that evaluation without requiring any user action — Cursor became a SpaceXAI product at closing. If procurement policies require using neutral tools or tools from approved vendors, SpaceXAI may not be an approved vendor, and the acquisition created a compliance gap retroactively.
The same question applies to any organization that has evaluated GitHub Copilot as a Microsoft product without fully accounting for the Microsoft-OpenAI relationship — both Copilot and OpenAI's API share infrastructure and product dependencies that make "neutral AI coding tool" a harder claim to sustain.
Thinking Machines Lab Inkling 2 Small
Thinking Machines Lab released Inkling 2 Small on November 5. The benchmarks:
Artificial Analysis Intelligence Index v5.0.1: 31 — 18 points above the first-generation Inkling Small, matching Google DeepMind's Gemini 3.8 Flash (high). If the Apache 2.0 weights ship next week, Artificial Analysis says this would be the highest-scoring open-weight model developed outside China. Vals AI called it "the highest-scoring US open-weight model by a wide margin."
SWE-Bench Pro v2 (Hard): 86.3% — exceptional on the hardest version of the standard software engineering benchmark.
FrontierCode 2.0: 25% (vs Claude Sonnet 5.5 at 11%) — strong on this coding-first benchmark.
Terminal-Bench 4.0: 13% (vs Sonnet 5.5 at 30%) — relatively weak on autonomous terminal task completion.
Architecture: 594B total / 17B active parameters, sparse MoE, native vision, variable reasoning effort. Trained on 12 trillion tokens in 68 hours. Post-trained with RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses.
The 68-hour training run is unusually fast for a model of this parameter count — it reflects the efficiency gains from MoE architecture (only 17B parameters active per token means training is effectively comparable to a dense 17B model per step, with capacity advantages from the sparse routing). This is the same design philosophy as Kimi K3.1 Pro, Muse Spark Open, and DeepSeek V4 — maximum knowledge capacity from minimum active compute.
Weights are closed at launch, pending external safety testing and red-teaming. The Apache 2.0 release is promised for next week. If it ships, it becomes the highest-scoring open-weight model developed outside China — a meaningful milestone for the Western open-weight ecosystem, which has been trailing the closed-weight frontier and the Chinese open-weight frontier simultaneously.
The name "Thinking Machines Lab" is new enough that limited institutional history is available. The lab appears to be a recent entrant with the technical expertise to execute a training run of this scale efficiently — 12T tokens in 68 hours on what must be substantial GPU infrastructure. If the weights ship next week with the promised Apache 2.0 license, the full model card and training details will make the lab's lineage clearer.
The coding tools arms race: what's left to acquire
The major AI coding tool acquisitions of 2026: - SpaceX acquired Cursor ($60B, $4B ARR) - Meta acquired Conductor (terms undisclosed) - Microsoft has GitHub Copilot natively - Anthropic has Claude Code natively - OpenAI has Codex and the new Agents API
What's left in the independent ecosystem: JetBrains AI (private, $6.5B+ valuation), Replit's Ghostwriter (independent, significant user base), Sourcegraph Cody (enterprise-focused, independent). These are the remaining neutral-ish distribution channels for AI coding capability that are not yet subsidiaries of a frontier model company.
The consolidation is not surprising — the value in AI coding tools is primarily the distribution to working developers, and every lab wants that distribution pointed at its own models. The question for enterprises is not whether this consolidation was predictable (it was) but whether procurement policies kept pace with the speed at which "neutral developer tools" became subsidiary products of competing AI companies.
