xAI releases Grok Code Fast 1 for agentic coding

Background

AI models have increasingly been used as coding agents, which read code, make edits and run tools over many steps. xAI's release gave coding a dedicated model of its own.^[1]

What happened

On 28 August 2025 xAI released Grok Code Fast 1, which it described as a speedy reasoning model for agentic coding.^[1] xAI said the model had been built from scratch with a new architecture.^[1] The model was generally available through the xAI API from launch.^[1]

Key details

On SWE-bench Verified, which tests whether a model can fix real issues in open-source projects, the model scored 70.8%. xAI ran this test with its internal harness, so the score comes from the company's own setup.^[1]

API pricing was $0.20 per million input tokens and $1.50 per million output tokens.^[1] For a coding agent that sends large amounts of code as input, the input price is a major part of the cost.

The model launched free for a limited time with GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode and Windsurf.^[1] Each of these is a coding tool in which a programmer works with an AI assistant. The free offer let developers try the model inside tools they already used.

xAI also said a version with a longer context was already in training.^[1] A longer context lets a model read more code in one request.

Agentic coding tools typically let a model read files, propose edits, run commands and check the results. A model built for this work needs to be quick enough that a developer does not wait long between steps, and cheap enough per token that many steps in one session stay affordable. Those two goals are the reason a fast, low-cost coding model was a distinct product choice.

Impact

The release showed a clear direction for xAI's model line: a fast, cheaper reasoning model made for agents, distributed through the tools developers already use. For readers, the main point is that the same model can be judged by a published benchmark score, a price per token and the list of tools that carry it, and each of these measures says something different about the model.