Grok 4.5 arrives with a focus on code and agents

xAI has launched Grok 4.5, an artificial intelligence model designed for programming, multi-step tasks, and working with documents, spreadsheets, and presentations. The company presents it as its most capable model to date.

The model was trained alongside Cursor, an AI-assisted programming tool, using data from code, science, engineering, and mathematics. It also went through hundreds of thousands of reinforcement learning tasks, a method that adjusts the model based on evaluations of its responses.

What Grok 4.5 can do

Its main goal is to handle complete technical projects, not just answer isolated questions. According to xAI, it can take on tasks such as:

• Write code in languages such as Rust, C, and C++.
• Create a functional application from a single instruction.
• Build complex Excel models using data researched on the web and formulas distributed across several sheets.
• Prepare PowerPoint presentations with native shapes and diagrams.
• Draft Word documents and leave notes to continue the work later.

In one of the examples shared, Grok 4.5 is instructed to create a simulation of the universe and the solar system with Three.js. The application must include orbits, realistic movement, stars, adjustable speed, and a modern interface.

The difference is that the model attempts to complete the project from start to finish. For you, that means it can move from a general description to a usable first version, although you will still need to review the code, the data, and the final result.

More speed and lower cost

Grok 4.5 processes up to 80 tokens per second, according to xAI. Tokens are the units the model uses to read and generate text, code, and other content.

The company says it needs approximately half as many steps as other leading models to solve comparable tasks. That amounts to nearly twice the token efficiency, although these are figures reported by xAI and depend on the type of task.

The announced price is:

• $2 per million input tokens.
• $6 per million generated tokens.

If these figures hold for long programming or analysis jobs, the cost may be easier to control than with models that need many interactions to reach the same result.

Where it is available

Grok 4.5 is already the default model in Grok Build, xAI's tool for creating applications and working with documents. It is also available in Cursor on all plans and in the SpaceXAI console through an API key.

The model was trained using tens of thousands of NVIDIA GB300 GPUs. xAI also highlights its data filtering and selection, including deduplication, quality evaluation, and the selection of specific content for each field.

For you, the important point is not just that Grok 4.5 writes code. It is that the model is designed to keep technical tasks going for longer, research information, create files, and complete multiple steps without asking for instructions at every turn. The question now is how reliably that autonomy works outside demonstrations and in real projects.