Cursor released Grok 4.5 with SpaceXAI and made it available across desktop, web, iOS, CLI and its SDK on the same day, turning the model from a research claim into a product users can try now. Cursor says the model is its most intelligent release yet, and it is being offered with significant usage in individual and team plans, plus doubled usage for the first week.
That timing matters because people are looking for a model that can do more than write code, and Cursor is presenting Grok 4.5 as the answer. The company says it is the first model it has built for more than software engineering, designed for difficult, long-running tasks that require using tools creatively to solve problems. It is priced at $2/M input tokens and $6/M output tokens for the base model, while the fast variant is $4/M input tokens and $18/M output tokens.
Cursor is also trying to show that this is not just a larger version of its earlier work. It said Composer 2.5 was its coding specialist, while Grok 4.5 was trained with a broader mix of data and reinforcement learning on hard problems in realistic environments. Those environments were built from trillions of tokens of Cursor data, including interactions with codebases and software tools, so the model could learn from existing software and developer-agent behavior as well as high-quality STEM tasks, research papers and other knowledge work.
The training setup is where the promise and the pressure meet. Cursor says many of the problems were made difficult enough that even frontier models fail at them, yet it still calls Grok 4.5 its most intelligent model. To build those environments at scale, it said engineers defined the problem and the way a solution would be checked, then large groups of agents constructed, tested and refined the environments, some of which would have taken hundreds of engineers months to build by hand. In other words, the model was trained to keep working when the easy answer is wrong.
That same capability also explains why Cursor added new safeguards around cybersecurity. The company says Grok 4.5 can be useful in software engineering, data science, finance, legal work and other computer-based tasks, but it is also signaling that the model’s reach creates new risks it has to manage. Composer 2.5 will remain available, and Cursor says it will keep releasing new models of that size going forward, leaving Grok 4.5 as both a next step and a test of how far the company wants to push beyond coding.
