Grok V9-Medium: 1.5 Trillion Parameters and Cursor Data to Chase Claude's Coding Crown

The model powering Grok right now runs on 0.5 trillion parameters. Its replacement triples that number. Grok V9-Medium, announced by Elon Musk on May 25, has finished training and targets a public release in mid-June 2026.

Grok V9-Medium: a 1.5T model built for code

Musk confirmed on X that xAI completed training on the model, which is three times bigger than the v8-small currently serving all Grok production traffic. Fine-tuning is underway, and reinforcement learning starts within days.

V9-Medium is not the flagship 6-trillion-parameter Grok 5. It's a separate model with a different focus: coding. When asked if it would handle code better, Musk's answer was blunt: "Much better at coding."

Cursor developer data fuels Grok V9-Medium training

Musk revealed that xAI incorporated extensive Cursor data during training. Cursor is the AI code editor used by developers at OpenAI, Stripe, and Perplexity. With over 4 million active developers, Cursor generates real engineering workflows: multi-file edits, large-codebase refactors, and correction cycles where developers accept, reject, or modify model suggestions.

SpaceXAI reached a deal with Anysphere (Cursor's parent) that includes an option to acquire the company for $60 billion later this year. That Cursor data pipeline is clearly a long-term integration play, not a one-off training decision.

Can Grok V9-Medium close the gap with Claude and ChatGPT?

Here's the uncomfortable truth. Grok holds roughly 6% enterprise AI adoption, compared with 55% for OpenAI, 47% for Anthropic, and 39% for Google. On SWE-bench Verified, Claude Opus leads at 87.6% while Grok 4 reaches 75%.

App downloads also tell a story: Grok fell from over 20 million in January to 8.3 million in April, a near-60% decline. The consumer momentum has stalled, and xAI badly needs a win on the developer front.

Grok V9-Medium's real bet: edit-heavy workflows

V9-Medium's advantage may live in edit-heavy developer workflows rather than from-scratch generation. If your workload is refactoring, review, and multi-file edits, it could be the more interesting test. That's a smart niche to target, given how Microsoft is building its own coding model for Copilot.

Raw parameters alone won't close a 12-point SWE-bench gap. But real developer workflow data from Cursor could be the kind of training signal that actually moves the needle. We'll know in a matter of days.