NEW Claude 3.7 Sonnet Is Simply The Best Coding AI (Claude Code Testing)

All right, ladies and gents, I want to show you something here. Claude 3.7 came out, it's probably the best LLM for code generation. And it came out with this thing called Claude Coder. This is some serious competition for a lot of the AI powered IDEs out there. All of them charge subscription fees. This thing is available just through claude. So rather than talk about it, let me just show you through this. I'm genuinely excited about this thing because. Because I just built this in like two minutes. And I'm not saying it's revolutionary that I built this in like two minutes. Okay. This has been possible for months now, okay. With since 3.5 Sonnet essentially building things like this is not that hard with the appropriate applications. But look at that. Expenses, income reports, savings, tips generated by AI. All of this was generated in the CLAUDE interface from a one shot prompt like this. Okay? And then I took all the code, dropped it into a folder, and rather than going through like the, I don't know, 30 step process of actually setting up all the files and splitting it all up, I just used Claude Coder here. Claude code, that's what it's called. It's a research preview. It came with this release. We'll talk about it in a second. And I just navigated to the folder and told it. Yo, bro, what did I tell it? You know, set this all up for me. I told it something like that and then it gave me an error message when I ran it and I just gave it the error message and the whole thing works. Okay. Before this, I spent about one hour trying to make it work manually and I was still not done. This is something that only apps like Cursor could have done up until now. And correct me if I'm wrong, I'm not a, you know, professional software engineer here, but I feel like I have a pretty good grasp on all these apps that are coming out. And you know, I'm a novice coder myself. Like I self taught myself Python like 3, 4 years ago, but essentially it just ran the whole thing and voila, here we are, working web app that is, you know, with authentication and everything else, look, I can log out, I can register with new accounts. Let's go through the step by step, no cuts, no edits. I want to show you how this works and I want to use Claude Coder to actually improve upon this application in the middle of this video. How about that? So I'm just going to actually start there. Okay. I'm going to open up this terminal so you can See how this Claude code thing works. And again, this is not revolutionary. Right? These things have been accessible through applications. Like, I don't know, what do we have, like Cursor or Windserve or Lovable now or Pythagora. All of these apps that are kind of a. Well, not really a copilot, like a full IDE with native AI integrated could do this. This comes for free. They just put this out there. It's a research preview. Okay, so to play around, you only need to install this. I'm not going to go through installation. Basically, if you've dabbled with code just a little bit, you'll manage to figure it out. They have great documentation on it over here on their blog. Yeah, you just go here on this link under Claude coder and then I'm sure you'll be able to figure it out. Maybe a little bit of AI help would help. Just help twice. But yeah, here, under joining this preview, here are all the steps you need Node, JS and Git installed. And then you basically run this and navigate to the folder that you want to be in. And then you just say claude and that's it. So once you have it installed, you. You literally just open up a terminal with the folder that you want to, that you want to work in and you just say claude and then, boom, Claude is working with the code. And then, point being, you can do things like this. You can go in here and say, my favorite prompt of all time, make it better. And. Or like, let's just say make it better would be good, but I don't want to mess up too much here. I'm just going to say make it prettier. And now, you know, Claude 3.7 sonnet is going to figure out what that means, implement all the changes, change my folder structure, whatever it has to do in all of these folders that it just created for me to make it prettier. And basically here for the terminal, I'm talking to it in human language and natural language rather than like prompting it and giving it code. Again, nothing revolutionary. What is revolutionary though, I suppose, is this new model. So while it does this, let's look at this new model. This is just going to keep working for a while. It's going to update my code base and we can look at the new version of the app in a second here. Okay, but what is brand new is this new model Claude 3.7 sonnet and Claude code. So I started with Claude code because I thought this was really sick, but Claude 3.7 Sonnet is their new model. If you've been following the space, you might have known that 3.5 sonnet was up until O1 Pro or like O1 basically considered the best model to generate code with. Right? It wasn't even close. Everybody actually using it and building apps was like, hey, all of them are great, but Claude 3.5 sonnet for code generation, that's where it's at. Now we have 3.7 sonnet and is much better now. I should say that like, you know, since the release of 3.5 sonnet, so people have been talking about 01 and 01 Pro and Deepseak R1 and then now O3 mini as superior in terms of code generation. Now there were still debates, right? It wasn't black and white. It wasn't like people kind of like turned on Claude and were like, oh, we don't use that anymore. Many people still, up until like yesterday were still using Claude 3.5 Sonnet to generate their code. Many of these apps were still running with it to write code like cursor, etc. But now we have just a way better version of the best model for code previously. So now it should be pretty clear. The benchmarks, they speak a clear language. If you Compare it to 3.5 Sonnet, well, on all of these benches it just beats it. Look at this. This is 01, this is 03 mini on the high setting. It gets 49.3% on software engineering bench. And this one with the custom scaffolding, which essentially I looked into this. Basically what this means is that they added some prompting and they let it think more than 30 times. By default it thinks up to 30 times. They let it think like up to 100 times. In most cases it just fought like 40, 50 times. But basically they let it reason more and they gave it some extra prompting to like store its answers and it reached 70% versus OpenAI's O3 mini high at 49.3%. Noted. It should be noted that this does not include Grok free because grokfree didn't publish this yet and their API is not out yet, so we can't run these benches ourselves. And it doesn't include O3 full because that isn't released yet. So fair enough. So compared to all the other options right now, killer model in terms of coding, okay? And this is all like coding focused, but I think this video is going to be also relevant for non technical folks. As you should know what's going on here. This is wild and maybe you can even use it for yourself. Agentic tool use. Look at that. Amazing. All these other benchmarks, I think this is the main one. It just wins on almost everything compared to some of the competition. And like, you know, Grok is like strong competition. And some things like visual reasoning, ooh, Grok has it. High school math competitions, ooh, Grok completely kills it over here, you know, but we don't have all the Grok benchmarks on things like graduate level reasoning, this actually comes out on top. Look at that. The top number with the scaffolding comes out on top. So legitimately, everybody loved 3.5 Sonnet already. This is a massive, like, this is the main comparison. I wouldn't even compare it to the others. I would Compare it to 3.5 Sonnet because it's not just about the benchmarks, it's also about, like, the vibes. It's about the taste test, if you so want. And on that, it was always killing it. Although some of these were far ahead in benchmarks, people were still using 3.5 sonnet and now 3.7 sonnet is just a ginormous upgrade, even in terms of benchmarks now. How does it perform in practice? Well, only time will really show. But what we can do today and what we can do now, instead of running all of these random prompts of like, generate me visualization of this dashboard of this, like, that's all well and good, I think there's value to that. But, like, what we're going to do now is we're going to actually build the application for ourselves. Actually, I did that in advance in preparation for this video. And we're going to improve the application with this thing they called Claude code. And Claude code, basically simple. I'm going to mute this, you can check out this video. But once you have it installed, which I don't know for people who have the very basics of coding, or like, if you know how to use a terminal, if you know how to run a few basic commands, like this shouldn't take you more than like three to four minutes. And once you have it, you basically just right click any folder and then you say services and you just open up a terminal, a window, and you just say Claude. Boom. That's it, right? And then once you have that, you can talk to it in natural language just like you would talk to me or fellow engineer, or just like you would talk to Cursor or one of the other apps. I keep throwing up my hands a lot because I'm like, this is big, but it Is So that's what Claude code is. Okay. Working with it, as simple as we pointed out. Maybe let me just round out this blog article so we can close it and leave this behind. You know, they're saying, like, basically up until now we had assistance in 2024. Now this is the first milestone where we actually have collaborators that we work together with, aka, you know, other companies name this, they call this agents. And then eventually we're going to have pioneers which like come up with breakthrough solutions by themselves. Other, you know, OpenAI calls this like AGI. So, yeah, there you go. This is the blog post. Very interesting. Let's look at what this thing does in practice and how it performs versus some of the competition here. And a few more details in the end, like the outputs are massive. Like it just gobbles up 20,000 words. Like it's nothing. Okay. The same prompts. I just want to say this upfront. The same prompts. I've been testing this between ChatGPT and Claude. You tell ChatGPT to write you a 50,000 word essay and it comes back with like a thousand words and it's like, well, this is all I could do. Claude created 21,000. I can show you the example conversation here if you want. Claude created 21,000 words and was like, hey, this is all I could do. But if you just type continue, I'll keep going. And then you type continue and it generates another 20,000 words for you. Insane. I was running some prompts here. As you can see, I was obsessing over it over the past few hours. We, by the way, just, we spent three hours in my office hours that I hold every second Monday right now in the community just playing with this thing. So it was really fun. But here it is. So check it out. It just wrote. It just wrote this. Okay, look at how long I have scrolled. It just wrote all of this in a one shot prompt. I believe. Correct me if I'm wrong, but I believe this is the longest output length web AI Assistant LLM that we have. If you use it for the API, by the way, you can output 128,000 tokens in one shot. That's another interesting thing here with like, which is like, wow. So if you use the anthropic console, you can just output, you know, even more. But this, as you can see, we'll just throw this into the tokenizer and then we'll move on to the app. Okay, but I just wanted to cover this point here. If we throw it into the tokenizer, this was the ChatGPT output, by the way? 1300 tokens, 7000 characters. And this right here, 20,000 tokens, 110,000 characters. My apologies. I said words. It's tokens. So I guess the tokens would be like, I don't know, 6, 16, 17K, if my napkin math here is right. Anyway, 110,000 characters in one shot versus what? These are the 01 Pro and ChatGPT. And these are the 01 Pro and ChatGpt 4O outputs. Okay, 7,000 characters versus 6,000 characters versus Sonnet 6. 7. 3.7 at 110,000 characters. Okay, so output length is insane, which means it can create these large code bases and create large apps for you in one shot. And then, you know, if. If that's not long enough, it tells you, like, hey, I can't write 500,000 characters. But just tell me. In the end, somebody told me, like, just tell me continue and I'll continue. Yeah, I just told it continue, basically, and then it kept going. There you go. So right here you said you can write continue to keep the chat going. Okay, so you get the point. Massive, massive output. Really good at coding. What does it look like in practice, though? Let's see. So it went ahead and did something here. What did it do? While I was talking about all this. Okay, it came up with a bunch of improvements. Shows you all the code it changed. The app now has a more professional, cohesive look. While maintaining full functionality, users will find the interface more engaging and intuitive. All right, so while I was blabbering away here, this upgraded the app. Let's see how that went. Can I just refresh here? Yeah. Look at the visual upgrade. Okay, so I'm gonna. Wait. I had a username. Was it Test one? I think it was Test one. No. Okay, so I'll just register a new account. How about that? Create a new account. Say test two. I'll give it some random password, create account. And there we are. How about that? That is a upgraded interface, right? So these are the types of things that you could only do with, like, Cursor up until now, or you had to piece it all together manually, which meant, like, updating every single file. It was basically a pain. That's why everybody's using these apps like Cursor and the competition that I kind of, like, named this. It's free though, right? I just installed this from their page that I showed you here. Like, you literally just install these two things, and then, like, you install it with this command, and then on your computer you can just say Claude. You could go into a folder, a thing you built and you just say Claude and then Claude does things for you. Pretty insane. Like, I mean, what am I supposed to say? We're doing this live sort of, right? No edits, just results. Insane. Look, more saving tips. And then I can enter various things. So I'm going to like, just briefly, no thinking involved. I'm just going to give it random numbers. Okay, test two random numbers. Maybe change the date. You can see all of this is working. Like this works as well or as better as code generation with Sonnet did. Whatever. Add some of those and then add some expenses. One, you have categories. This is amazing. It's created a personalized expense tracker app for me with login management with a database in the background. Okay, this might be too much. Whatever, it doesn't matter, you get the point. Okay, so add all of those and then if I look at the reports, voila. Okay, obviously this one is way above, but you get the point, right? And then it does like forecasting for you and everything and where you'll be with your savings rate and everything. Obviously these numbers are like a mess now, but you get the point. You get a beautiful dashboard, you get your top expenses and then based on those you have saving tips. Now I think the one thing that didn't work is this AI implementation. But again I could easily tell it to, hey, like implement the Anthropic API. I would have to mess with the API key, which I don't want to do right now here. But you get the point, like if something doesn't work, you just follow up here and it just works. And this is free, there's no subscription. And for our pic, just put this out and this model is so damn good and outputs something so long that some of these other apps, I'm not saying they're redundant, like there's certainly a lot of uses for them, but like if you just want to build something quickly and just try something out and if you don't want to pay a subscription fee, then there you go. Cloth code. Pretty, pretty awesome. Okay, so I think that's a pretty neat demo. What else could we do? We could like follow up to do random stuff. I mean, I don't want to make this video too long, but I want to show you the process that I created this with because it's as simple as could be, right? It's this one. I prompted it to build a personal finance tracker using Python and web based dashboard. Users should be able to log expenses and now, hold up. We're going to compare this to zero one Pro and Grok. Okay? So users should be able to log expenses, categorize spending, and visualize trends with matplotlib. This app should include an AI feature that suggests ways to save money based on spending patterns. So how about this just for fun, I'm going to follow up with. The AI feature doesn't generate new recommendations. And what I'm going to do, I mean, let's just follow up. It should probably ask me for my API key, which I'm not going to have here, but that's fine. I think you get the point. It just generated all of this in one shot. Okay. It gave me one file, which was yay long. Impressive. I just downloaded this file, threw it in a folder, and with CLAUDE code, I told it set all of this up for me, and it created the file structure. It did everything. And when I hit play, when I started, basically it didn't work. It actually gave me an error, and then I just copied the error that I got in my browser into CLAUDE code, hit Enter, and it was like, ah, sorry, fixed that for you. And then it worked. So it was like the simplest process, just like you're used to. But the point is, this thing works. This thing works. Okay. Basically here I was trying to set it up manually before I was like, okay, what folder structure do I need? And then it tells you you need to run all of this, create all of these folders. Let's look at the comparison. What does this look like inside of. Yeah, here was generating some prompts to run. So that's where I have this from. I was running a deep research, but basically if I Compare it to 01 Pro, 01 Pro things for 3 minutes and comes up with, yeah, like, okay, you need this project structure. You need to set this up. Okay, cool. Then you need to set up a database. Then you need to set up your Flask app. Then you need to do this, you need to do that. Step number seven, you can. You need to put it all together, right? By the time I've done this, I'll build you a house with my bare hands, and I'll start a family with multiple children. By the time you put all of this together manually, as a, as a coding novice, like, this is ridiculous. Nobody's going to be doing this. I mean, I guess you can, but, like, come on. The point of AI is supposed to be it's easy, right? Not complicated. Grok. What did Grok tell us? Grok basically goes in and is like, hey, here's all the requirements, okay, that you asked for. Cool, cool, cool. And then again we're like, here's the implementation plan. Install all of these packages, right? So it doesn't give you the commands. It's like, you got to install these, bro. And you got to know what I mean by that. This is the project structure I want from you. Then start Flask. Then, you know, do all of this again, like step number 2745, implement the visualizations, implement these, you know, fetch the expenses for the current. Like, embed it in the dashboard. I mean, don't get me wrong, Grok is amazing. I've actually been using it over the weekend and like, hats off, like, holy shit, great model. But this takes a lot of efforts to implement. Whereas Cloud 3.7 was just like, yo, here's everything. And then inside of CLAUDE code, I just like, navigate to the folder and tell it go. So let's see, what is it doing here? This implementation provides diverse, helpful financial recommendations. So did all of this as I was talking again, right? Added variety, implemented refresh functionality, add the advanced analysis. Didn't even want an API key. How does that work? I don't know. I'm going to refresh down. Let's see how our saving tips. It was testing to, right? So you know, not to. Was it test to. Not to yap on too long here. Let's round this video out. But the saving tips, let's see. This is. Work better. Yeah. Now they refresh. There's more variety. Like, it improved it. It did it, right? It's that simple. And you don't need a subscription. You can just run this. It's amazing. Okay, so I see. I think you see the point. We looked at the comparison, we looked at the token output, we looked at CLAUDE code, we looked at the model itself and how good it is at generating stuff. By the way, I want to show you one more example, which is kind of a funny one. I think this is sort of, as Chris in the community pointed out, it's sort of an Easter egg when you ask it. One of these standard problems, like, you know how many Rs are in Strawberry? It's like, click the strawberry to find out and it creates this. And then you're like, huh, what is going on here? And they're like, free. So this is what you get from cloud 3.7 versus, you know, I don't know. If I run this inside of Zero1 Pro, it will probably be a bit drier. So the combination of, like Anthropics 3.7 and they're just like the taste test, the vibes of like 3.5 already. These benchmark results, the fact that you have Claude code now and you can like circumvent some of these paid applications that cost like serious money, all of that makes this release. This is AI for the people. This is, this is what you actually want to be using to build things. So I encourage you to try this yourself. Go in, run some prompts that you want, check out the blog article. This is amazing. I'm impressed. I'm gonna be. I know what I'm gonna spend all week doing. I'm gonna be building things like this personal finance tracker and more with their free application. And we're gonna be doing the same thing inside of the community. If you want to learn more about that, we have technical people who can help you set this up, get this going, get tips and tricks, get your questions answered. We are generally speaking non technical focused community for people who are not software engineers and want to do this stuff. So we come at it from that angle. If you want help, check out the community. If you want free like develop the free development assistant, check out Claude code. My mind is still blown by this. And with that being said, I hope you enjoyed this overview and I hope you're going to be getting a lot of value from this. The AI wars are happening, everybody's releasing stuff and Claude just killed it on the coding front. Love it. That's all I got for today and I'll see you soon.