Grok 4.5 Arrives: xAI and Cursor Drop a New Frontier Model

xAI and Cursor jointly release Grok 4.5, a mixture-of-experts model trained on trillions of tokens that extends beyond coding into broader knowledge work — priced at $2/M input tokens.

Grok 4.5 launched yesterday as a joint release between xAI (SpaceXAI) and Cursor, marking the company's first model intentionally built for more than software engineering. Elon Musk described it as an "Opus-class model."

The mixture-of-experts architecture was trained on trillions of tokens, including a massive dataset of Cursor user interactions — capturing how developers work and how AI agents interact with their environments. Unlike the coding-specialist Composer 2.5, Grok 4.5 was trained on a deliberately broader mix of STEM tasks, research papers, and knowledge work.

Cursor used reinforcement learning on deliberately difficult problems to train the model. These environments were constructed at scale by distributed agent systems — some would have taken teams of hundreds of engineers months to build manually. The approach pushes the model to investigate problems, use tools creatively, recover from mistakes, and verify its own results.

Pricing lands at $2 per million input tokens and $6 per million output tokens for the base model. A fast variant runs at $4 per million input and $18 per million output. Cursor subscribers get double usage for the first week.

The model is available today across desktop, web, iOS, CLI, and SDK. Composer 2.5 remains offered alongside it, with Cursor committing to support both model weight classes going forward.

Sources: Cursor Blog, TechCrunch