On June 16, SpaceX agreed to acquire Anysphere — the company behind Cursor — for $60 billion in stock. It is the largest acquisition in AI developer tooling history. Twelve days later, before the deal even closed, Cursor session data was flowing into supplemental training for Grok 4.5. The model launched publicly on July 9. If you have been coding in Cursor over the past several months, that work is now baked into a model owned by Elon Musk’s rocket company. Here is what changed, what it means, and what to do.
Your Sessions Were the Training Data #
This is not a metadata play. Cursor’s structured session data — the exact prompt you typed, the suggestion the model produced, the edit you accepted, the error the compiler caught next — is the most valuable training signal that exists for a coding model. At four million active seats, that feedback loop is worth more than any parameter count. xAI knows this. The Grok 4.5 launch page says the model was “trained alongside Cursor.” The company also disclosed that “an earlier Cursor codebase snapshot was accidentally included in training.” Read that sentence twice.
The privacy shift happened quietly. Cursor’s standard Privacy Mode — the setting most users leave untouched — now permits session data to flow into Grok 4.5 training. This changed on June 28, before the acquisition closed. If you did not opt out, you did not opt in either. You were simply enrolled.
Right now, before you read another paragraph: Open Cursor, navigate to Settings, go to Privacy and Data, and verify your training inclusion status. If you work with proprietary code, you may want to switch to API-key mode and route requests through Claude or GPT instead of Grok.
Grok 4.5: What the Data Built #
Grok 4.5 launched July 9 on all Cursor plans. It is now the default model in Grok Build and available across the xAI API, SpaceXAI console, and Microsoft Office add-ins. The performance numbers are credible: 64.7% resolve rate on SWE-Bench Pro, a 500K context window, and roughly 80 tokens per second. SpaceX claims 4.2x token efficiency versus Claude Opus 4.8 on the same benchmark — 15,954 output tokens per task against 67,020. Pricing is $2 input/$6 output per million tokens, with $0.50 cached. That undercuts Opus 4.8 significantly.
Where the numbers get murkier: Grok 4.5 beats Opus 4.8 on DeepSWE 1.0 and Terminal-Bench 2.1 but loses on DeepSWE 1.1 and SWE-Bench Pro head-to-head. SpaceX’s internal benchmarks carry an obvious conflict of interest given that Cursor data gave Grok an advantage on coding evaluations. Independent verification is still sparse as of this writing. The model is capable — but whether it outperforms Claude Sonnet or GPT for your actual workflow is a question worth testing with your own tasks, not trusting the launch deck.
One clarification worth making: you are not forced to use Grok. Cursor still accepts API keys from Anthropic, OpenAI, and others. Grok is the new default — it is not the only option.
Independent AI Coding Tools No Longer Exist #
The Cursor deal completes a consolidation that has been running all year. GitHub Copilot belongs to Microsoft. Windsurf was acquired by OpenAI in March for $3 billion. Google shut down the open-source Gemini CLI on June 18 and replaced it with a closed binary. Now Cursor belongs to SpaceX. The list of independent, privacy-neutral AI coding tools above a garage-project threshold is effectively empty.
This is not catastrophizing — it is the market structure. When you choose an AI coding tool, you are choosing which conglomerate receives your code, your prompts, and your debugging patterns. The choice used to be about features. It is increasingly a vendor-concentration decision.
What to Do Now #
Four concrete steps, in priority order:
Check your Privacy settings in Cursor today. Settings → Privacy and Data → verify training inclusion. Do not assume your previous settings survived the June 28 policy change.Use API-key mode for sensitive work. Cursor lets you route requests to Claude or GPT via your own API key. This bypasses Grok and the associated data pipeline entirely.Enterprise teams: audit your vendor contracts. The zero-data-retention policy that cleared your legal review may have quietly changed. Get updated Data Processing Agreements from Anysphere before the Q3 close.Test an open-source fallback.Aider(Apache 2.0), OpenCode (MIT, 165k+ GitHub stars), and Continue.dev (VS Code and JetBrains) all run locally against any OpenAI-compatible API. Having one configured is cheap insurance.
The acquisition does not close until Q3 2026. That is not a comfort — Grok 4.5 is already live and the data pipeline is already running. The window for auditing your setup is now, not after the paperwork is signed. Enterprise teams in regulated industries should treat this as an active compliance issue, not a future concern.