xAI completes training of 1.5T-parameter Grok V9-Medium

xAI said it has completed training of Grok V9-Medium, a 1.5-trillion-parameter foundation model, with a public release planned for mid-June. The headline detail is that it was trained on Cursor developer workflows — a deliberate bid to make Grok strong at agentic, real-world coding rather than just benchmarks.

The model represents roughly a 3x parameter jump over the current Grok generation, signaling xAI is scaling aggressively. Training on Cursor-style workflows suggests xAI is optimizing for the multi-step, tool-using coding patterns that distinguish capable agents from autocomplete.

The framing is explicitly competitive: it's positioned as xAI's answer to Claude's coding dominance, arriving the same week Anthropic touted 80% Claude-authored code and OpenAI shipped GPT-5.3-Codex. Coding is now the contested frontier, and every major lab is racing to claim it.

The skeptical note, surfaced in community reaction, is that Grok V9-Medium reads as 'reactive catch-up' rather than a leap ahead, and a finished-training announcement is not a shipped, benchmarked product. The mid-June release — and independent evals against Opus 4.8 and GPT-5.3-Codex — will determine whether the 3x parameter bump translates into a real coding edge.