Grok Code Fast 1

// Overview

Grok Code Fast 1 is xAI's first model built specifically for agentic coding, released on August 28, 2025. Internally codenamed "sonic," it is positioned as a speedy and economical reasoning model that handles the high-volume, iterative work of software development — reading code, calling tools, and editing files inside coding agents — rather than a general-purpose chat model.

xAI says it built Grok Code Fast 1 from scratch on a new architecture, using a pre-training corpus rich in programming content and post-training on datasets reflecting real-world pull requests and coding tasks. The model is text-only (no image, audio, video, or document input), ships with a 256K-token context window, and was tuned for low latency and aggressive prompt caching so it stays interactive during agent loops. xAI reported throughput of roughly 90–100 tokens per second in partner workflows.

At launch the model was offered free for a limited time through GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf, then priced very cheaply on the xAI API at $0.20 per million input tokens and $1.50 per million output tokens. Grok Code Fast 1 was deprecated on May 15, 2026, with retirement scheduled for August 15, 2026, as xAI moved its coding line toward newer Grok Code / Grok Build models.

// Benchmarks

1. SWE-bench Verified70.8%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Offered free for a limited time at launch through partners including GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf.

Pricing source ↗

// Strengths

• Purpose-built for agentic coding: reading, editing, and tool-calling inside coding agents
• Very low cost — $0.20 input / $1.50 output per million tokens
• Low latency with high throughput (~90–100 tokens/sec reported), keeping agent loops interactive
• 256K-token context window for working across larger codebases
• Aggressive prompt caching ($0.02/1M cached input) suited to repeated agent prompts
• Strong reported coding result: 70.8% on SWE-bench Verified (xAI's internal harness)

// Best for

• Agentic coding inside tools like Cursor, Cline, GitHub Copilot, and Windsurf
• High-volume, iterative code editing and refactoring where cost and speed matter
• Codebase navigation and multi-file changes within a 256K context
• Drafting and resolving pull-request-style fixes
• Cost-sensitive coding pipelines that call an LLM many times per task

// How to access

// Grok Code — every version

The full lineage of the Grok Code line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.