Grok Code Fast 1 in 8 Minutes

--- type: transcript date: 2025-09-02 youtube_id: SoWr_K09w4Y --- # Transcript: Grok Code Fast 1 in 8 Minutes XAI has just released their much anticipated coding model GRO code fast one. In this video, I'm going to go over the blog post and show you a practical demonstration of the model within cursor. They describe the motivation between this model as a lot of other models while they're powerful, they don't often feel purpose-built for agentic coding workflows where loops for reasoning as well as tool calls can often times feel frustratingly slow. And because of that, the engineers at XAI as heavy users of agentic tools themselves, their engineers saw room for more nimble, responsive solutions optimized for day-to-day tasks. So from that, this is the model that they built. Now, in terms of the pre-training, they used a large rich corpus of programming related content. But for post- training, what they did is they curated highquality data sets that reflected real world pull requests and coding tasks. And that's a common criticism of a lot of models is oftent times they do perform well on certain benchmarks but don't actually perform well in real world tasks. Next up, if you are watching this video shortly after recording, you will be able to access this for free for a limited time within GitHub Copilot, Cursor Klein, Rue Code, Kilo Code, Open Code, as well as Windsurf. In terms of the pricing for the models, it costs 20 cents per million tokens of input, a $1.50 per million tokens of output, and only 2 cents per million tokens of cashed input. If we take a look at this model when compared to some of the other flagship models that are out there, whether it's Gemini 2.5 Pro, GPD5, Claude 4 or Gro 4, we see that this model's throughput as well as price is considerably cheaper than some of these other generalpurpose models in terms of benchmark. So on Swebench, they did score a 70.8, which is up there with some of the best results that are out there. But one thing that they mentioned is while SWEBench provides valuable insights, that doesn't exactly tell the full story. They described that we found that this doesn't fully reflect the nuances of real world software engineering, particularly the enduser experience in agentic coding workflows. There is some praise in here from some of the key stakeholders within the agentic coding community where a lot of people describe this model as being both fast as well as accurate and a good overall model for agentic coding. So within here over the next few weeks, you can expect rapid iteration on this model and if you have any feedback, you can go ahead and provide this to the team over at XAI. in terms of testing the model. So what I've done is just spun up a Nex.js application and we're going to go through and work through some of the motions and see how it performs. The first thing that I want to test is its design sensibilities. So I'm going to say create a modern SAS landing page. And right off the bat, what we see here is we see the model reasoning through it creates a plan based on our request. So it's broken out all of those tasks to create that SAS landing page into eight various tasks. So first up, it's creating a hero section and iteratively it's going through step by step. What it's doing here is it's going through and creating all of the relevant components. So if I take a look at the hero component, the future component, the pricing, it is a very fast model. We see a number of these components are within the hundreds of lines of code. And here is what it has generated for us. So we have all of the different components that it's created. We also have this nice frame motion animation effect all throughout the pages. We have the pricing, we have the frequently asked questions, love by teams. If I just scroll through the page here, we have a reasonable starting point for our application. There are a couple things that I don't love like the linear gradients. But one thing that I do like is it actually didn't use any emojis. Instead, what it did as one of the first step is it did install an icon library and overall it does have a pretty good structure in terms of a starting point for an application. Now, if I take a look at the code, so we are using client components for all of these. We do have those animation effects. We do still have our server rendered page. We have the navigation. we have the respective components all throughout all of this. Now, to test out the edit functionality. So, I'm going to say I want to remove all of the linear gradients and instead I want the overall look and feel to be much brighter. So, I want a modern white as well as black feel for all of what we've created so far. Okay. So, I'm going to go ahead and send that in. It's created a to-do list. And what it's doing now is it's methodically going through line by line changing out all of the respective pieces of what we need to update within our codebase. As soon as this done, I'll circle back and we'll take a look at what it's generated. And here we go. So, here is what it has generated for us. So, within here, we have this overall modern look and feel with these brighter whites instead of the linear gradients that we had on some of the elements previously. If I just scroll through the page here, it's basically updated all of the respective pieces of our application. Next, I want to test it with a bit more of an involved problem. I'm going to say I want to create a page for 3JS that has an interactive environment of a cube that I can interact with. Additionally, I want to create a dashboard page that's rich with visualizations as well as leverage a really good charting library that has a variety of different charts. We see it planning. We see it thinking through and all of the different responses that you have within cursor here. From what I found with a model that does have this token speed that is around 200 tokens per second, it actually is quite hard to follow unless you actually break the cursor and read through step by step what it's doing. because all of the different planning stages you'll see that it will just flash there for a second. But in terms of the UX, it will be interesting to see as these models get faster how this actually evolves over time. For instance, do we need to see all the intermediate thinking steps within something like cursor by default? Maybe we don't need to actually see that and see all of this jumping around. Maybe there can be a little bit more of a graceful UI for something like this. Before I move on, I want to thank today's sponsor, brilliant.org. Brilliant takes a fresh approach to learning that perfectly complements these AI developments. Imagine a fitness center for your mind where every lesson makes you mentally stronger. Instead of sitting back, you're diving in, tackling challenging problems with crystal clear explanations. Research shows this approach is six times more effective than conventional learning methods. Their latest course on large language models especially caught my attention, offering a comprehensive yet digestible exploration of AI technology. You don't need to be a coding wizard or an academic scholar to benefit from Brilliant. It's designed for anyone eager to enhance their analytical thinking and run their understanding of modern concepts. Whether you're on a daily commute, lunch break, or evening downtime, Brilliant transforms spare moments into valuable learning opportunities. To learn for free on Brilliant, go to brilliant.org/developers digest. Scan the QR code on screen or click the link in the description. And Brilliant's also given our viewers 20% off an annual premium subscription which gives you unlimited daily access to everything on Brilliant. So I'll take a look at 3JS here. So here is the interactive cube that it's created for us. We have this text. We also have on hover it create this red cube and on click it changes sizes. So while it isn't much the one core piece of this is it was able to leverage that library correctly in one shot. This is a starting point. Obviously, you could test this with some more involved prompts. Maybe it wouldn't quite do as well. But what I find when you're leveraging different libraries is if you start with something relatively simple and iterate from there within your agent coding process, you actually can get quite far with whatever you're looking to build. Next up, in terms of the dashboard page, in one shot, here is the page that it created for us. We have these line charts. We have these interactive bar charts as well as pie charts. And with just one shot, it was able to decompose that 3JS task as well as this dashboard task into two functional prototypes for a starting point within our application. Now, obviously, the overall look and feel definitely does need some work, but if we did initially focus a little bit more on the design, this overall look and feel would probably be quite a bit better. All right, but otherwise, that's pretty much it for this video. Kudos to the team at XAI for what they put out. This is definitely both a very fast and performant model. And I am curious, let me know what you like about the model, don't like about the model. Are you going to be using this over something like the Gemini series of models, GBT5, or the models from Anthropic? Let me know in the comments below. If you found this video useful, please comment, share, and subscribe. Otherwise, until the next one.