xAI Finishes Training 1.5T-Parameter Grok V9 on Cursor Coding Data

Elon Musk announced that xAI has finished training Grok V9 at 1.5 trillion parameters — roughly 3x the size of the current production Grok — with heavy use of Cursor programming data as a deliberate bet on coding capability. Musk said the model ships publicly in 2-3 weeks and is a stepping stone toward Grok 5.

The Cursor-data angle is the strategic move: where OpenAI relied on broad code corpora and Anthropic invested in agentic post-training, xAI is leaning explicitly on real developer interaction traces from a tool that ranks among the most-used AI IDEs. Combined with Tesla FSD telemetry and the live X firehose, community commentary frames Grok V9 as a 'data moat' play more than a pure scale play.

The same week, Greg Brockman called GPT-5.5 'a uniquely good coding model,' Alibaba's Qwen3.7-Max debuted at #4 on Code Arena 'on par with Claude Opus 4.6,' and Anthropic shipped Opus 4.8 at 88.6% SWE-bench. The coding leaderboard is the most contested benchmark of Q2 2026, and Grok V9 enters that race directly.

Skeptical takes: HN commenters questioned the AGI framing Musk attached, and worried that Cursor data raises consent and competitive questions for Cursor itself. Watch for the Code Arena placement once V9 is live.