From Zero to 76.5 Billion: The Grok Code Fast Review After Token Surge
We partnered with xAI and announced that its users can use Grok-code-fast-1 ( the stealth model you might also know as “Sonic”) for free within the Kilo Code extension without any rate limits. This drove unprecedented token usage by early adopters , which spearheaded us to 76.5 billion tokens, as of August 29th - just 3 days after the model launch.
Our users are true early adopters, showing 10x the usage compared to other extensions. Let's review what makes this model special, so you can jump on board and enjoy free frontier AI access
TL;DR: Grok-code-fast-1 is great at grep, terminal commands and file editing and you get free access to use it inside Kilo Code.
Grok-code-fast 1 at a glance
Grok-code-fast-1 is the newest frontier thinking model from xAI. Built from scratch with a brand-new architecture, this reasoning model addresses the speed bottlenecks that plague traditional coding assistants during iterative development cycles.
Key Specifications:
Context Window: 262k tokens
Speed: 190 tokens per second
Performance: 70.8% on SWE-Bench, 90% cache hit rates
Pricing: $0.20 per million input tokens, $1.50 per million output tokens
At $0.20 per million input tokens and $1.50 per million output tokens market rate, you can do the simple math and see that even with a simple project (Aviation Instrument Panel) where I used 24.6k tokens, I saved (~4.9k): $0.001 on input tokens and (~19.7k): $0.0295 on output tokens, for a total cost of about $0.03. While 3 cents might seem small for this project, the savings scale significantly:
100 similar projects: $3.05 saved
Daily coding (if you used 100k tokens/day): $12.50 saved per day
Monthly heavy usage: $375+ saved
Stress testing the model
Following xAI's official prompt engineering guide, I created an "Aviation Instrument 6-Pack" simulation to stress-test their claims about Grok-code-fast 1 excelling at tasks such as file editing, grep, and terminal interactions.
Create a vertical speed indicator, altimeter, turn coordinator, earth speed indicator with marked stall speed, turn coordinator that should have a plane icon, magnetic north indicator and the artificial horizon.
The initial output was impressive, but I pushed further to test file editing capabilities with complex requirements: maintain existing class inheritance hierarchy, preserve public/private method signatures, and enhance the turn coordinator with slip/skid indicators and rate of turn functionality.
File Editing Performance
I enhanced the turn coordinator with a slip/skid indicator while preserving existing instrument panel architecture—executed flawlessly without breaking other components. Grok-code-fast-1 did exceptionally well while executing precision modifications and did not break any other components. I asked the model to change my codebase into TypeScript, and it understands language-specific patterns and maintains proper formatting:
Terminal interactions
Initially the project was written in JavaScript and I asked the model to rewrite its entire codebase into TypeScript. It maintained TypeScript conventions and handled terminal interactions well. For instance, it detected that I do not have pnpm installed inside the root directory or globally, and it maintained the TypeScript /dist file structure. The UI looked the same; you can preview the TypeScript code here.
How to use the model in Kilo Code
Install/Update the Kilo Code VS Code extension to version 4.83.1 or later
Settings → Providers → Select Kilo Code → 'x-ai/grok-code-fast-1'
Start coding with unlimited frontier AI
Bottom Line
Grok Code Fast delivers on its core promises: exceptional speed and tool mastery. For detailed discussion of pros and cons I would recommend this Reddit thread. In my experiments I noticed limitations such as architectural thinking (additional nudging needed for specific implementations, otherwise model defaults to simple vanilla JavaScript ) and domain expertise (needed to provide extra prompts to explain what parameter each avionic instrument includes), its strength at the moment is predominantly speed, as Kilo Code community members pointed out:
The 76.5 billion tokens used by our community in just 3 days speaks volumes about both the model's capabilities and our users' enthusiasm. With completely free access through Kilo Code, there's no reason not to try it.