Grok Code Fast 1

What it is

Grok Code Fast 1 is a fast reasoning model for agentic coding, released by xAI on 28 August 2025 ^[1][2]. Agentic coding means the model is meant to work inside a coding assistant that plans and carries out multi-step changes across a codebase, rather than only finishing the line in front of it. Work of this kind involves many consecutive calls, so response speed matters as much as quality. It is a coding-specific model: xAI placed it alongside its general Grok 4 line rather than as a replacement for it ^[2].

How it was built

xAI said the model was built from scratch on a brand-new model architecture ^[1]. Its pre-training corpus was rich in programming content, and its post-training data was curated to reflect real-world pull requests and coding tasks ^[1]. xAI also said it refined the model together with its launch partners so that it behaved well inside their agentic platforms ^[1].

What it is good at

xAI said the model was particularly adept at TypeScript, Python, Java, Rust, C++ and Go ^[1]. On SWE-bench Verified — a benchmark built from real software issues, in which a model has to produce a change that resolves the issue — it scored 70.8% on the full subset using xAI's internal harness ^[1]. A score of this kind depends on the harness and the subset used, so the figure is best read as xAI's own measurement.

Price and caching

At launch the model cost $0.20 per million input tokens, $1.50 per million output tokens and $0.02 per million cached input tokens ^[1]. The last figure refers to prompt caching: parts of a prompt that repeat across requests can be stored and reused instead of being processed again, which lowers both cost and latency. xAI reported prompt-caching optimisations that regularly achieved cache hit rates above 90% with those partners ^[1].

How it was distributed

At launch it was available in GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode and Windsurf, where it was free for a limited time ^[1]. It was also generally available through the xAI API from launch, and xAI said a variant with extended context length was in training ^[1]. Context length is the amount of text — code, files, conversation — a model can take in at once; a longer context lets an assistant keep more of a project in view.

Where it sits in the line

Grok Code Fast 1 is xAI's coding-specific model, sitting alongside the general Grok 4 line ^[2]. The later Grok Build tool, released in 2026, continued this line of coding-focused work ^[2].