Cursor's Grok 4.5: A General-Purpose Model for Text and Business Work

Cursor, together with SpaceXAI, has released Grok 4.5, a general-purpose text model aimed not only at programming but also at data science, finance, legal work and other knowledge-work tasks.

What Grok 4.5 can do

The new model was built as a universal tool for complex, long-running tasks. Unlike its predecessor Composer 2.5, which focused exclusively on code, Grok 4.5 was trained on a broader dataset: STEM tasks, scientific papers and materials from various professional fields.

• Architecture: mixture-of-experts, co-developed by Cursor and SpaceXAI
• Training: trillions of tokens of data on how users interact with codebases
• Use cases: programming, data analysis, finance, legal work
• Base price: $2 per 1 million input tokens and $6 per 1 million output tokens
• Fast variant: $4 per 1 million input tokens and $18 per 1 million output tokens
• Availability: desktop, web, iOS, CLI and SDK

How Cursor trained its smartest model

The key element of training was reinforcement learning on tasks in realistic environments, covering both software development and broader intellectual work. The environments were deliberately made hard enough that even frontier models failed to solve them.

To scale the process, Cursor built a distributed agent system: engineers define a task and the criteria for checking a solution, and groups of agents automatically create, test and refine the training environments. According to the company, some of these environments would have taken teams of hundreds of people several months of manual work.

Cursor also openly disclosed an important caveat: Grok 4.5 showed an advantage on the internal CursorBench benchmark because an early snapshot of Cursor's codebase accidentally ended up in the training data. The exact scale of the effect is unknown. The company has already removed this data for future models and is working on an updated benchmark.

Pricing and access

Grok 4.5 is already available in all Cursor versions: desktop, web, iOS, CLI and SDK. Subscribers to individual and team plans get a significant amount of model usage within a pool, and during the first week the company doubles that limit. Composer 2.5 remains on offer, and Cursor plans to release new models of this class in the future.

Alongside the release, the company added new safeguards that take Grok 4.5's cybersecurity capabilities into account. Their details are not disclosed, but the fact that they were included suggests the model is seen as a potentially powerful tool in this area too.

What this means

Cursor is betting on a shift from narrowly specialized coding assistants to general-purpose agents for knowledge work. Grok 4.5 is the company's first attempt to move beyond the developer niche and offer a tool that works equally well with code, legal documents and financial data.

Frequently asked questions

How does Grok 4.5 differ from Composer 2.5?

Composer 2.5 was specialized only in programming. Grok 4.5 was trained on a broader dataset (STEM tasks, scientific papers, knowledge work) and is designed for data analysis, finance and legal work, not just development. Cursor plans to develop both classes of models in parallel.

How much does it cost to use Grok 4.5 via the API?

The base version costs $2 per 1 million input tokens and $6 per 1 million output tokens; the fast variant costs $4 per 1 million input tokens and $18 per 1 million output tokens. Subscribers to Cursor's individual and team plans get a significant amount of usage at no extra charge, within the pool of primary models.

How does Grok 4.5 differ from the previous version?

The new Grok 4.5 was trained on a broader dataset (STEM tasks, scientific papers, professional materials), unlike its predecessor Composer 2.5, which was specialized exclusively in code.

What is Grok 4.5 for?

For data science, finance, legal work and other knowledge-work tasks, extending its capabilities beyond programming.