Elon Musk confirmed on June 5 that Grok V9-Medium has completed training. The specs: 1.5 trillion parameters, approximately three times the current production model (V9-small at roughly 500B parameters). Evaluation results described as positive. Public release targeting mid-June. Currently rolling into Tesla's connected-car fleet and the X social network.

Benchmark numbers: not published yet.

The interesting thing about this model isn't the scale. It's what the model was trained on, and who gets to see it first.

The Cursor data pipeline

V9-Medium was trained on Cursor data — actual developer workflow data from one of the most-used AI coding tools. Not curated training sets. Not synthetic code. Real sessions: debugging, refactoring, navigating unfamiliar codebases, knowing when to stop and ask. The messy middle of software work, not just the clean completion tasks that coding benchmarks test.

This matters for a specific reason. Most coding benchmarks — SWE-bench Verified included — evaluate isolated task completion. Fix this bug. Implement this function. Real developer work is iterative: read ten files, change one, make sure you didn't break something three layers away. Training on Cursor workflow data means V9-Medium has seen this pattern at scale.

The Cursor connection isn't incidental. SpaceX — xAI's primary compute infrastructure parent — holds a $60 billion acquisition option on Anysphere, Cursor's developer. That option becomes exercisable after SpaceX's June 12 IPO. Whether SpaceX actually acquires Cursor or not, the data pipeline is already built into V9-Medium's training run.

The distribution argument

Here's the part that doesn't show up in any benchmark.

When OpenAI ships a model, users get it through ChatGPT or the API. When Anthropic ships one, same mechanism. When xAI ships one, they push it via OTA update to six million internet-connected Tesla vehicles and surface it to hundreds of millions of X users. Instantly. Without anyone downloading anything.

That six-million-vehicle fleet collectively generates the equivalent of over 100 years of real-world driving scenarios every single day, per xAI's own reporting. Sensor data. Voice interactions. Real-world model feedback that flows back into training. Other AI labs have to buy or license training data from third parties. xAI has a fleet that produces it continuously.

X adds the social-conversation data layer. Tesla adds the real-world sensing layer. Cursor adds the developer-workflow layer. No other AI company has this trifecta of proprietary streams.

The flywheel logic: bigger deployment → more real-world signal → better next training run → better model → bigger deployment. It's the same loop that made Google Search so hard to compete with fifteen years ago. Musk is building the AI version of it.

Source spread

• TechTimes — V9-Medium training complete — [hype]. Musk's June 5 confirmation. Good on scale but light on benchmark details.
• TechTimes — Grok V9 distribution flywheel — [builder]. Best available breakdown of the Tesla + X deployment strategy.
• KuCoin flash — Cursor data integration — [builder]. Confirms the Cursor workflow data pipeline detail.
• ChatForest — Builder's guide to V9-Medium — [skeptic]. Good on the V9-small vs. V9-Medium distinction and why they're not interchangeable.

Pros & cons

What's real:

• The 3× parameter scale-up is a real bet. It might not translate linearly into capability, but it's not a marketing number — xAI burned compute to do this, and that cost is real.
• Cursor workflow data as a training signal is a sound hypothesis. Real developer sessions cover problem-solving patterns that curated datasets miss. If the data collection was done well, this is a genuine differentiator.
• The OTA deployment model is genuinely novel. No other AI company can push to 6 million connected vehicles and a major social platform simultaneously. That's a different kind of scale than cloud API adoption.

What deserves a side-eye:

• No public benchmarks. "Positive evaluation results" from the developer is not independent verification. SWE-bench Verified, TAU-bench, MMLU — we need the canonical numbers before making any production decisions. Don't skip this step.
• The Cursor data provenance is unclear. Were Cursor users informed their workflow data would train Grok V9-Medium? CNBC flagged this concern in April. Worth understanding before deploying V9-Medium in enterprise settings with confidential codebases.
• Tesla driving data is real-world but it's driving data. It's a rich multimodal stream for autonomous systems; it doesn't obviously improve TypeScript quality.

What builders need to know

Further reading

• TechTimes — Grok V9 Rolls Into Tesla and X — distribution strategy
• TechTimes — V9-Medium training complete — training confirmation
• TechCrunch — SpaceX / Cursor acquisition option — data pipeline backstory
• ChatForest — Builder's guide to V9-Medium — pre-release technical overview
• CNBC — Promises and risks of AI chatbots in cars — the data provenance concern