AI Deck

Grok Code — xAI's fast, low-cost model built for agentic coding (retired in May 2026)

Grok Code (officially grok-code-fast-1) is an AI model released by xAI on August 28, 2025, purpose-built for agentic coding work. It was trained from scratch, separately from the general-purpose Grok 4 line, with post-training centered on real pull requests so that it handles the full loop of reading code, fixing it, and running tests. Its defining traits were a generation speed of roughly 190 tokens per second and an unusually low price of $0.20 per million input tokens, and it was rolled into major coding tools all at once — GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode, and Windsurf.

As of August 2026, however, this is no longer a model you can newly choose. xAI retired eight models including Grok Code Fast 1 on May 15, 2026, and API requests addressed to grok-code-fast-1 are now automatically routed to its successor, grok-build-0.1. GitHub Copilot ended support on the same date. This article covers what the model was and what to pick today for the same kind of work.

Key Features

  • Built entirely around speed: Response speed of roughly 190 tokens per second. In coding work where an agent calls tools dozens of times, the wait per call translates directly into perceived speed. It was positioned as a model that moves quickly rather than one that deliberates
  • Optimized for tool use: Trained around the operations a coding agent uses daily — searching code with grep, running the terminal, editing files. The design targets multi-step automated work rather than one-off question answering
  • Pricing built around prompt caching: $0.20 input / $1.50 output / $0.02 cached input, all per million tokens. For agent workloads that re-read the same codebase context repeatedly, xAI reported cache hit rates above 90 percent with launch partners
  • A clearly stated set of strong languages: xAI stated strength in TypeScript, Python, Java, Rust, C++, and Go, covering the major languages used in practice from web applications to systems programming
  • Simultaneous integration across major tools: At launch it shipped in GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode, and Windsurf, with free access for a limited period in some of them. A SWE-Bench-Verified score of 70.8% was published

Pricing

ItemPrice (per million tokens)Notes
Input (while offered)$0.20Official price at the August 2025 launch
Output (while offered)$1.50Same as above
Cached input (while offered)$0.02Applied to repeatedly read context
Current (August 2026)grok-code-fast-1 is retired. Requests are routed to grok-build-0.1, and the official migration guide states billing at $1.25 input / $2.50 output

Pricing is current as of August 2026. Check the official site and the model list for the latest rates.

Pros & Cons

Pros

  • Trained specifically for agentic coding, so it fits automated work that calls tools repeatedly
  • Fast responses, keeping an agent’s iteration cycle short
  • While it was offered, it was strikingly cheap for a coding model, making heavy token usage cost-predictable
  • Extremely low-cost cached input favored workflows that work the same repository over and over
  • Broad integration across major coding tools kept switching effort low

⚠️ Cons

  • Retired on May 15, 2026, so it can no longer be newly selected
  • The API model identifier still resolves, but the responses come from the successor model, so neither the original behavior nor the original pricing is reproduced
  • Because it was tuned for speed, general-purpose higher-tier models suit deliberate design decisions and harder reasoning better
  • Its strong languages were stated explicitly, but performance on other languages and niche frameworks was never publicly documented
  • It is an API model rather than a standalone product, so use depends on support from the editor or agent side

Comparison with Similar Services

ItemGrok Code Fast 1Grok Build 0.1Grok 4.6GitHub Copilot alternatives
ProviderxAIxAIxAIOpenAI / Anthropic
PositioningSpeed-focused coding modelSuccessor agentic coding modelGeneral-purpose flagship (recommended for coding)Recommended migration targets in Copilot
Context length256K tokens256K tokens500K tokens
AvailabilityRetired (May 15, 2026)AvailableAvailableAvailable
Main useHigh-iteration automated code fixesAgentic coding and web developmentHarder design work and long-running agentsEveryday coding assistance inside Copilot

GitHub points Copilot users to GPT-5 mini and Claude Haiku 4.5 as replacements, while xAI officially recommends migrating to grok-build-0.1 for direct API use.

Who Is It For

  • Developers who embedded grok-code-fast-1 in a tool or script and now need to choose a migration target
  • Anyone looking for a cheap, fast model to hold down the running cost of a coding agent
  • Users who want to stay on xAI models but have lost track of which one is meant for coding
  • People who follow model generational shifts using public benchmarks such as SWE-Bench

Summary

Grok Code Fast 1 embodied the idea that for agentic coding, speed and low cost sometimes matter more than raw intelligence. Its 190 tokens per second, $0.20 input pricing, and cache-oriented billing design meaningfully lowered the operating cost of coding agents. That role has since passed to its successor, grok-build-0.1, and the model itself was retired on May 15, 2026. For the same work today, grok-build-0.1 is the straightforward choice on the xAI API, with Grok 4.6 for harder tasks. If you still have old configuration in place, it is worth checking that you are not being billed unintentionally on a different model at a different rate.

← Blog