Introducing Grok 4.5

Summary

Cursor and SpaceXAI have released Grok 4.5, a mixture-of-experts model positioned as their most capable system to date. Unlike the earlier Composer 2.5, which was trained as a coding specialist, Grok 4.5 draws on a deliberately wider data mix that includes high-quality STEM tasks, research papers, and other knowledge-work sources alongside trillions of tokens of Cursor interaction data. This combination is intended to support both software engineering and tasks in domains such as data science, finance, and legal work.

The model was trained with reinforcement learning on problems set in realistic environments that require tool use, recovery from errors, and result verification. These environments were generated at scale through a distributed agent system in which engineers define a problem and verification criteria while agents construct, test, and refine the setups. The approach allows the creation of tasks difficult enough to remain informative even as frontier models improve.

Grok 4.5 is available immediately across Cursor’s desktop, web, iOS, CLI, and SDK interfaces. Individual and team subscriptions include substantial usage quotas, doubled for the first week, while the base model is priced at $2 per million input tokens and $6 per million output tokens; a faster variant carries higher rates. Composer 2.5 continues to be offered alongside the new model.

On CursorBench the model shows an advantage traceable to an earlier snapshot of the Cursor codebase that was inadvertently included in training data. That data has since been removed for subsequent models, and the benchmark itself is undergoing a larger revision.

Why it matters

This update is highly relevant for product teams and builders as it introduces a powerful new AI model integrated directly into development workflows via Cursor. It offers actionable pricing, SDK access, and expanded capabilities beyond standard coding tasks, enabling Dutch AI practitioners to accelerate product development.