xAI launches Grok Code Fast 1 for agentic coding

xAI has launched grok-code-fast-1, an artificial intelligence model designed to write and modify code inside tools such as Cursor, GitHub Copilot, and Windsurf. Its main focus is not just response quality, but reducing the time developers spend waiting.

The model is now generally available through the xAI API. It is also free for a limited time on several partner platforms, including GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode, and Windsurf.

Designed to work inside a project

grok-code-fast-1 was not created as a standard question-and-answer assistant. xAI trained it from scratch to work with coding agents, systems that can read files, run commands, search text, and edit code on their own.

The model is prepared to use common tools such as grep, the terminal, and file editors. In practice, that allows it to handle tasks such as:

• Creating a project from scratch.
• Answering questions about an existing codebase.
• Fixing specific bugs without modifying unnecessary parts.
• Implementing features according to a step-by-step plan.

According to xAI, it is particularly capable in TypeScript, Python, Java, Rust, C++, and Go. The company says it can complete many common tasks with little supervision, although that does not eliminate the need to review changes before adding them to a real project.

Speed is the main difference

xAI says it has optimized both the model and the infrastructure running it to speed up every response. In agentic coding workflows, the system typically alternates between reasoning, calling a tool, and analyzing the result. Every wait adds up.

The company says grok-code-fast-1 can make dozens of tool calls in the time it takes a user to read the beginning of its explanation. It also says the model achieves cache hit rates above 90% with its launch partners. Caching allows parts of an already processed request to be reused, reducing both response time and cost.

That speed changes how you work. Instead of sending one huge instruction and waiting for the agent to do everything, it may be more effective to break the work into smaller tasks: first plan a function, then implement it, and finally fix any errors that appear. That way, you can review and redirect the model at every step.

API results and pricing

On the full version of SWE-Bench Verified, a set of real-world programming problems, grok-code-fast-1 scored 70.8% in xAI's internal tests. The company warns that this type of benchmark does not fully reflect the day-to-day experience of using an agent inside an editor.

That is why xAI also combines automated tests with programmer evaluations and reviews of the model's behavior on common tasks. Speed, reliability, and ease of correcting course matter just as much as the ability to solve an isolated problem.

The API pricing is:

• $0.20 per million input tokens.
• $1.50 per million output tokens.
• $0.02 per million input tokens reused from the cache.

A token is a small unit of text that the model processes. The final cost will depend on how much code it reads, how much it generates, and how many times it repeats the project context.

What comes next

xAI initially released the model quietly under the codename sonic and published new versions to respond to community feedback. The company promises frequent updates, with improvements arriving in days rather than weeks.

It also says it is already training a variant with multimodal input, parallel tool calls, and a larger context window. That would allow it to analyze more types of information, perform several actions at the same time, and work with larger projects.

For you, the news matters if you code with agents inside an IDE: you can try a model focused on reducing wait times and offering low API costs. The key will be seeing whether that speed preserves accuracy when tasks become long, ambiguous, or affect many parts of the code.