Grok 4.5: SpaceXAI and Cursor build a model from agent data

SpaceXAI released Grok 4.5 on July 8, a mixture-of-experts model trained jointly with Cursor, the AI coding tool. The model is not a general-purpose chatbot. It is built for “agentic software, engineering, and workflow tasks,” according to SpaceXAI’s documentation. At $2 per million input tokens and $6 per million output tokens, Grok 4.5 sits in a competitive pricing tier alongside models from OpenAI and Anthropic. But the training process is what makes this release unusual.

Cursor published the technical details alongside the launch. The training included “trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools.” That dataset is not synthetic. It is real developer behavior recorded across thousands of codebases, capturing how humans and agents interact with development environments. Cursor trained the model jointly with SpaceXAI, using a deliberately broad data mix that included STEM tasks, research papers, and knowledge work beyond pure software engineering.

The model uses a 500,000-token context window, supports function calling and structured outputs, and is available in Cursor across desktop, web, iOS, CLI, and the Cursor SDK. SpaceXAI also offers a fast variant at $4 per million input tokens and $18 per million output tokens.

Training on agent behavior

The key innovation is the training data. Previous coding models, including Cursor’s own Composer 2.5, were trained primarily on static code repositories. Grok 4.5 trains on the dynamic process of software development: how a developer navigates a codebase, what tools they call, how they recover from errors, and how they verify results.

Cursor used reinforcement learning on “difficult problems in realistic environments spanning both software engineering and broader knowledge work.” These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results. The company notes that “many of these problems had to be designed to be difficult enough that even frontier models fail at them.”

To construct these environments at scale, Cursor built a distributed agent system. Engineers specify a problem and a verification method. Large groups of agents then construct, test, and refine each environment. Cursor claims that “some would have taken teams of hundreds of engineers months to build.” This is a recursive training pipeline: the previous generation of models helps generate training data for the next generation.

Pricing and positioning

At $2 per million input tokens, Grok 4.5 is cheaper than OpenAI’s o3 ($10 per million input tokens) and comparable to Anthropic’s Claude Sonnet 4 ($3 per million input tokens). The $6 per million output tokens is competitive for a frontier coding model. Cursor includes significant usage in individual and team subscription plans, with double usage for the first week.

The pricing signals a market where coding agents are becoming a commodity input, not a premium service. SpaceXAI is betting that volume and integration with Cursor’s ecosystem will drive adoption, not exclusive capability. The model is also available via API in us-east-1 and us-west-2 regions, with rate limits of 150 requests per second and 50 million tokens per minute.

The CursorBench admission

Cursor disclosed one notable limitation. Grok 4.5 has an advantage on CursorBench, the company’s internal benchmark, because “an earlier snapshot of the Cursor codebase was accidentally included in training.” Cursor states that “the exact impact is unclear” and that the data has been removed for future models. The company is working on a larger update to CursorBench.

This is a standard data contamination issue, but it is unusual to see a company disclose it proactively. The admission suggests confidence that the model’s real-world performance will hold up against third-party evaluation. On SWE-Bench Pro and Terminal-Bench, Grok 4.5 shows self-reported scores above third-party models, though Cursor notes that the GPT 5.5 score on SWE-Bench multilingual comes from an internal run.

What this means for AI builders

Grok 4.5 represents a shift in how frontier coding models are built. The training data is no longer static code. It is the process of coding itself. Cursor’s distributed agent system for generating training environments points toward a future where model improvement becomes increasingly automated: the model helps build the environments that train the next model.

The joint training arrangement between SpaceXAI and Cursor is also notable. SpaceXAI provides the compute infrastructure, trained in its Memphis data centers, and the model architecture. Cursor provides the training data and the application layer. This is a partnership model that other AI labs and application companies may replicate: the infrastructure provider and the data-rich application company co-owning the model.

For builders, Grok 4.5 is available today at competitive prices. The real question is whether the training on Cursor interaction data generalizes to non-Cursor workflows. The model is designed for agentic tasks across software engineering, data science, finance, and legal work. If the training data is too specific to Cursor’s interface and tool set, the model may not transfer to other environments. Cursor’s deliberate inclusion of broader knowledge work data suggests the company is aware of this risk.

SpaceXAI and Cursor have released a model that is cheaper than many competitors, trained on a novel dataset of real agent behavior, and backed by a recursive training pipeline that may accelerate future improvements. The data contamination disclosure is a rare moment of transparency in a field where benchmark scores are often treated as marketing material. Grok 4.5 will be judged not by its CursorBench score, but by whether it makes agents more reliable in production.