SpaceXAI (formerly xAI) unveils Grok 4.5, built for coding and agents

SpaceXAI (formerly xAI) has unveiled Grok 4.5, a new large language model that it calls its most powerful to date. The model is aimed primarily at developers, engineers, and anyone who works on technical problems.

Who Grok 4.5 is for

According to the company, it built the model for writing code, running agentic workflows, and handling everyday knowledge work. That sets Grok 4.5 apart from general-purpose models, which try to handle creative writing, everyday questions, and programming equally well and often trade quality in each area for versatility.

• Developer: SpaceXAI, the company formerly known as xAI
• The model is called Grok 4.5, and the company positions it as the most powerful in its current lineup
• SpaceXAI trained it together with the team behind the Cursor development environment
• According to SpaceXAI, the model uses roughly half as many tokens as competing models
• The model focuses mainly on coding, agentic workflows, and technical work

Why the Cursor partnership matters

Cursor is one of the best-known development environments with a built-in AI assistant, and programmers around the world use it. Training the model with the Cursor team means SpaceXAI tuned Grok 4.5 for real programming workflows from the start, rather than adapting a general model to them after release. For developers, this approach usually translates into a sharper grasp of code context, fewer bad suggestions, and smoother behavior inside the development environment itself, where the model was tested during training.

What halving token usage means

SpaceXAI's claim that Grok 4.5 uses roughly half as many tokens as competitors matters most for agentic workflows, where the model carries out a chain of actions on its own: it reads files, calls tools, edits code, checks the result, and repeats the loop when needed. In these scenarios, the model can make dozens or hundreds of calls in a row, and token costs quickly add up to a significant bill for any team building a product on top of it. If the model really is twice as efficient as its competitors, that directly cuts the cost of running agents built on Grok 4.5.

Where the market for coding models is heading

Just a couple of years ago, most labs shipped general-purpose models and only later adapted them for code through fine-tuning and dedicated modes. SpaceXAI took a different route: it built Grok 4.5 around developer tasks and agentic workflows from the outset instead of retrofitting a general-purpose model. Partnering with Cursor during training is part of the same strategy. Rather than releasing a model "in a vacuum" and waiting for the programming community to fit it into their tools, the company builds the model into a workflow developers already use.

This approach also reflects a shift in how the industry makes money. Agentic workflows, where AI independently runs long chains of actions, don't need a one-off answer to a question; they need stable, predictable, cheap performance across hundreds of steps. That's why SpaceXAI puts as much emphasis on token efficiency as on the model's capabilities: for anyone building products on Grok 4.5, the two are tightly linked.

What it means

SpaceXAI's bet on developers, code, and agentic workflows shows where competition between AI labs is heading: the winner isn't only the model that is "smarter" in a general sense, but the one that runs cheaper and faster in programmers' real work loops. Training the model together with the Cursor team signals that major AI labs are increasingly partnering with the makers of tools that millions of programmers already use, instead of competing with them directly for users.