Grok Code Fast 1: xAI's Speed-Tuned Coding Model, Audited

xAI quietly shipped a coding-focused model called Grok Code Fast 1 on August 28, 2025¹, and the headline pitch is straightforward: the model is built for low latency and low cost in IDE-style agentic coding loops, not for topping every benchmark chart. It's worth understanding what it actually does well — and where the marketing leans harder than the data supports.

What Grok Code Fast 1 Is

Grok Code Fast 1 is a coding-focused reasoning model released by xAI on August 28, 2025¹. Before its public launch, xAI tested it under the codename "Sonic" and iterated on multiple checkpoints based on feedback from developer communities¹.

xAI says the model was built from scratch with a new architecture, a pre-training corpus heavy on programming-related content, and post-training data drawn from real pull requests and coding tasks². xAI has not officially published the parameter count in its launch announcement; third-party sources describe it as a Mixture-of-Experts design³, but treat that figure as unconfirmed by xAI.

Verified Quick Facts

Spec	Value	Source
Release date	August 28, 2025	xAI announcement1
Pricing (input)	$0.20 / 1M tokens	xAI2
Pricing (output)	$1.50 / 1M tokens	xAI2
Pricing (cached input)	$0.02 / 1M tokens	xAI2
Context window	256K tokens	xAI docs4
SWE-Bench Verified	70.8% (xAI internal harness)	xAI2
Output speed	~92 tok/s (xAI), ~106 tok/s (Artificial Analysis)	xAI2, Artificial Analysis5
Strong languages	TypeScript, Python, Java, Rust, C++, Go	xAI2

Speed: Fast, but Not "Almost Double GPT-5"

xAI cites approximately 92 tokens per second for the user-perceived loop². Independent measurement from Artificial Analysis puts API output at around 106 tok/s — above the median for reasoning models in its tier (~105 tok/s)⁵.

For comparison, GPT-5 (high) measures around 79.9 tok/s on its standard API⁶. That makes Grok Code Fast 1 roughly 15–30% faster than GPT-5 by output rate — meaningfully faster, but not "almost double." Earlier marketing copy floating that 2x figure does not match the published numbers.

Pricing: Genuinely Cheap, Not Free

The xAI API pricing is $0.20 per million input tokens, $1.50 per million output tokens, and $0.02 per million cached input tokens². That is among the lowest pricing for any frontier-positioned coding model — but "practically free" is a stretch. For high-volume agentic loops with large output bursts, the output token cost still dominates.

The "free" angle is real but limited. At launch, xAI partnered with GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf to offer the model at no cost for a promotional window¹. Free periods varied — GitHub Copilot's preview ran through early September 2025, Cursor offered a one-week free trial, and Windsurf Pro and Teams accounts had unlimited usage during the promo⁷. As of late 2025, Grok Code Fast 1 is generally available in GitHub Copilot under standard plans rather than free⁸, so the "free" claim is no longer broadly accurate.

Benchmarks: Solid for the Tier, Not Top-of-Chart

On SWE-Bench Verified, xAI reports 70.8% using its own internal evaluation harness². Independent third-party reproductions have come in lower — around 57.6% on some leaderboards⁹ — which is the typical gap when vendor harnesses with custom tooling are compared against standardized public runs.

For perspective, leading models on the SWE-Bench Verified public leaderboard sit well above this number in 2026. Grok Code Fast 1 is a fast, cheap coding model — not the top-of-chart agentic coder. That positioning is consistent with what xAI claims; problems arise only when secondary write-ups inflate it into "best-in-class."

Tool Use and IDE Fit

xAI's blog says the model has been trained specifically for tool calls common in coding agents — grep, terminal commands, and file editing — and that's where its IDE-friendly framing comes from². In practice, this means it can navigate a repository, edit multiple files, and run shell commands as part of an agent loop, which is what makes it useful in Cursor, Copilot, Cline, and the other launch partners¹.

What Developers Have Said

Verified developer reactions are mixed-positive. Shubham Saboo, an AI product manager and frequent open-source author, posted that Grok Code Fast 1 "is incredibly good at following instructions" and that it "competes with Claude Sonnet 4 and GPT-5 while being almost 10x cheaper"¹⁰. He demoed a Three.js interactive cube test, not a "3D crystal visualization" — small detail, but worth keeping accurate.

Mario Rodriguez, GitHub's chief product officer, said in launch coverage: "In early testing, Grok Code Fast has shown both its speed and quality in agentic coding tasks"¹¹. Earlier drafts of this post quoted a "Paul Kuver" calling it the fastest he'd tested; that quote could not be verified against any primary source and has been removed.

Limits Worth Knowing

A few honest limits to keep in mind:

• It's a fast/cheap tier, not a frontier reasoning tier. For deep architectural design, multi-step planning across very large codebases, or hairy debugging, Claude Opus or top-tier GPT-5 variants generally still come out ahead on independent leaderboards.
• Vendor benchmark vs. public benchmark. xAI's 70.8% SWE-Bench Verified number uses an internal harness; independent runs land lower. Treat both as data points, not gospel.
• Promo pricing ≠ list pricing. Many of the early "free on Cursor / free on Copilot" reactions came from promo windows that have since ended.
• Parameter count is unconfirmed. Third-party write-ups cite a 314B MoE design, but xAI's own announcement does not publish a parameter count.

Where Grok Code Fast 1 Actually Fits

If your workflow is high-frequency agentic coding — repository navigation, small-to-medium edits, terminal commands inside an IDE — Grok Code Fast 1's combination of low latency and low per-token cost is genuinely attractive. For large-output agentic loops where you'd otherwise spend significantly more on Claude Sonnet 4 or GPT-5 per call, the cost difference compounds quickly.

If you need top-of-chart reasoning, plan generation, or the highest possible SWE-Bench performance, this isn't the model for that — and xAI doesn't really claim it is.

TL;DR

Grok Code Fast 1 is a real release with real strengths: fast output, cheap tokens, decent SWE-Bench Verified numbers in xAI's harness, and broad IDE partner integration at launch. It is not a revolution, not "almost double" GPT-5's speed, and not still uniformly free. Use it where its tradeoffs match your workload, and skip the hype framing.

---

References

Footnotes

1. xAI, "Grok Code Fast 1." Official launch announcement, August 28, 2025. https://x.ai/news/grok-code-fast-1 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. xAI, "Grok Code Fast 1 Model Card," August 26, 2025. https://data.x.ai/2025-08-26-grok-code-fast-1-model-card.pdf ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

3. Third-party model summaries describing a Mixture-of-Experts architecture. Note: xAI's own announcement does not officially publish a parameter count. ↩

4. xAI Docs, "Models and Pricing." https://docs.x.ai/developers/models ↩

5. Artificial Analysis, "Grok Code Fast 1 — Intelligence, Performance & Price Analysis." https://artificialanalysis.ai/models/grok-code-fast-1 ↩ ↩²

6. Artificial Analysis, "GPT-5 (high) — Performance & Price Analysis." https://artificialanalysis.ai/models/gpt-5 ↩

7. GitHub Changelog, "Grok Code Fast 1 is rolling out in public preview for GitHub Copilot," August 26, 2025. https://github.blog/changelog/2025-08-26-grok-code-fast-1-is-rolling-out-in-public-preview-for-github-copilot/ ↩

8. GitHub Changelog, "Grok Code Fast 1 is now generally available in GitHub Copilot," October 16, 2025. https://github.blog/changelog/2025-10-16-grok-code-fast-1-is-now-generally-available-in-github-copilot/ ↩

9. BenchLM.ai, "SWE-bench Verified Benchmark — third-party leaderboard." https://benchlm.ai/benchmarks/sweVerified ↩

10. Shubham Saboo, public post on Threads about Grok Code Fast 1's instruction-following and pricing. https://www.threads.com/@saboo_shubham_/post/DN3NpLI2LeS ↩

11. Mario Rodriguez (GitHub CPO), quoted in launch coverage of Grok Code Fast 1. ↩