Grok 4.5 Is Live: xAI and Cursor's First Jointly Trained Coding Model Hits Opus-Class Performance at $2/MTok

TL;DR: Grok 4.5 is xAI's first Cursor-trained coding model: Opus-class performance, 1M-token context, $2/$6 per million tokens — live today in Cursor and the SpaceXAI API.

xAI launched Grok 4.5 on July 8, its strongest model to date and its first co-built with Cursor since the acquisition. The 1.5-trillion-parameter V9 foundation was fine-tuned on Cursor's engineering codebase and real-world coding interaction data — meaning it was trained on the same kinds of multi-file refactor, spec-to-code, and agentic debugging tasks that developers actually run in IDEs. According to xAI, it exceeds comparable frontier models on real engineering tasks with both intelligent and efficient reasoning.

Access: available now in Grok Build, in Cursor on all plans (no upsell), and from the SpaceXAI API console. EU availability expected mid-July due to data regulation review.

Pricing: $2/MTok input, $6/MTok output — undercutting Sonnet-tier pricing while claiming Opus-class performance. For teams routing agentic coding tasks through OpenRouter or a custom API client, Grok 4.5 is a legitimate benchmark target.

Why it matters: The Cursor acquisition gave xAI training data that's unusually well-matched to how developers actually use code models. A $2 input price at Opus-class quality changes the economics for agentic loops that currently use Opus. Run your task distribution against it before committing.

Sources: Introducing Grok 4.5 — SpaceXAI · Grok 4.5: xAI and Cursor's First Jointly Trained Coding Model — MindStudio · SpaceXAI launches Grok 4.5 — Axios