DeepSeekV4 + Claude Code = 100X Cheaper

[0:00]Imagine if you could use Claude code 100 times cheaper. In this video, I'm going to show you exactly how to combine the Claude code framework with the brand new DeepSeek V4.

[0:11]So you can save thousands of dollars, build as much as you want with no rate limitations.

[0:15]And with this brand new repo, set it up in minutes, even if you're starting as a complete beginner. And if you don't know who I am, my name is Jack Roberts.

[0:25]I built and sold my last tech startup, that was 60,000 customers. Now I'm building my own AI companies and I share the stuff here that I actually work. So, if you haven't already, grab that beautiful coffee and let's dive straight in.

[0:35]So, Claude for 100 times cheaper, let's go into it. And by the way, I've also tested this myself extensively, as well looked to all the benchmarks for you so I can tell you what the tank cleanser is, as well as my experience and when to use what for which, which you're going to find really helpful.

[0:48]So let's have a look at what DeepSeek V4 is, 1.6 trillion parameters, not something you can probably run on your own computer, put it that way, 1 million contact window and it's MIT open weights, which means it's fully compatible with the Claude code ecosystem.

[1:03]And one of the biggest benefits of DeepSeek is its tool calling ability. So all the tool calling we do in Tropic, we can now do a DeepSeek.

[1:10]That's what got me really interested in this originally is not just another connection, but it's something really cool. So, same workflow, significantly cheaper.

[1:19]So you look at this, Claude Opus is $62.50 a session, and this analogy, DeepSeek V4 can be significantly cheaper.

[1:27]Now, one thing to call out right now is they've got a promo on, which meant it's even more cheaper than usual.

[1:32]After May 5th, it's still significantly cheaper, but less than it is at the moment this week because we're running a special promo. DeepSeek V4 Flash is even crazily more cheap, basically 20 cents for 4-hour session compared to uh 62.50 for Claude Opus 4.7.

[1:47]All right, cool. So does it actually code? This is what got a lot of people excited about its benchmarks.

[1:52]So, SWE benchmark verified, you can see that is basically DeepSeek V4. How it compares to Sonet 4.6, how it's uh basically um Opus 4.7.

[2:03]You got Terminal Bench, and the MCP Atlas. I know what you're thinking, Jack, shoots great on the Tinder profile, but how does DeepSeek actually show up when we take it on a date?

[2:10]What does it actually look like? Because we don't want it just to be a benchmark queen, we want it to actually work in practice. But I just say this to say that it's it's hitting benchmarks and its baseline performance is good. But as always, we've really got to check out what this physically looks like.

[2:22]So, what's want to call out one quick thing here, which is this, is that the reason why this is a better strategy and the system I'm going to show in this video, by the way, works if you want to bring in DeepSeek, or if you want to bring in a completely free model, you can do that.

[2:34]But one of the big unlocks, even if you want to go 100% free, which by the way is an option, is that we have this dual terminal, which means that we can have a free model in one window.

[2:45]Uh we can have DeepSeek in another one, and then Claude code in a third, which means that if you wanted to do some work with Claude code and then tag in DeepSeek to crush it or tag in a free model, we can do all of this simultaneously, using the strategy that I'm going to show you inside this video.

[2:58]Now, to use DeepSeek, there are two ways we can do this. One is the OpenRouter, but I found you get rate limited so much, it's basically unusable.

[3:09]So I'm going to show a strategy of how we can use it with no rate limits as much as you want. So we want to head over to platform.deepseek and create a brand new account.

[3:17]And if you want to do this completely for free, I'll also show you that strategy as well. So, go ahead and create a brand new account. And here, you want to throw in a couple of dollars, could be $2, could be $5.

[3:23]Now, you can do this for free using OpenRouter and a series of things, that basically cycles free models, but again, it like get rate limited and it's not great.

[3:33]And free is not free because actually the quality's less, you get charged more. On DeepSeek, throwing like $2, $5, you can start using it.

[3:38]Then on the left hand side, come over to API keys, go ahead and copy this API key.

[3:43]Then the next thing I'm going to do is come over to this GitHub repo right here. This is a beautiful little repo. It's trending quite strongly right now.

[3:51]But the idea is it helps us use what I call that dual terminal split up, which means we can use the Claude code harness with anything that we want to. So all you do is come down to code, click on copy.

[4:00]Then head over to your IDE of choice. I'm going to be using anti-gravity for this one, so let's pull them up. Okay, wonderful. I am in anti-gravity.

[4:06]If this sounds like I'm speaking Spanish, check out this anti-gravity master class, I'll take you from zero to singing anti-gravity in your sleep.

[4:11]Now, all you're going to do is come up here and say, hey there, I'd like you to install and clone this repo and then what we're going to do is connect this to DeepSeek directly with the below API key.

[4:21]Now, the really cool thing we can do here, okay, is we can basically either connect to DeepSeek or we can connect to OpenRouter by grabbing our OpenRouter API key.

[4:30]And on openrouter.com, you come over, you literally can grab your API key from anywhere here. And the cool thing here is this obviously enables you to access any of the models on the entire planet that you want to.

[4:40]Um, so basically this applies for anything. And again, this enables us to use Claude on multiple different browsers at the same time.

[4:47]So now it's going to come down and grab that. What we want to do is grab the API key from DeepSeek. And as you can see, the repo is cloned and I can see exactly how it works.

[4:54]It's a proxy server that intercepts Claude API calls and routes them to alternative providers, including DeepSeek natively.

[4:59]The setup is simple, just copy - create a .env file with your DeepSeek key set as DEEPSEEK_API_KEY, configure the model routing (MODEL_TARGET) in your .env to point at DeepSeek, install dependencies via pip, and get the server running. So go ahead and paste your DeepSeek API key! Fantastic. That's now complete. Also, then ask it this, which is I want you to create two shortcuts. One is to run normally and the other one is to dangerously skip permissions, if you want to give it the ability to do that.

[5:11]But it's just an option if you want to. You may not want to be too ham-hock with it, with it being a different model. But it's there for you if you want to. So, here's everything we've done.

[5:18]We've installed it, it's all ready to rock and roll and we've got all the keys here, which is fantastic. And this just runs through everything it's got and our API keys are stored in the .env side over there on the left hand side. Cool, so it's actually tested working now.

[5:27]So to do this, click on terminal, click on new terminal and now it's up here. We're just going to type in DS for DeepSeek and then Claude should appear, which is cool.

[5:36]And then just going to say something like, hey, I don't know, who is the tallest man to ever live, something like that.

[5:41]Now, a quick hack, by the way, if you ask this model, who is it, they will basically tell you it is Claude Sonet 4.6. That's because they're reading the data that sits in it.

[5:51]That's not the actual model. So now we've basically fully connected to DeepSeek. What we could now do, if we wanted to, if I opened a brand new terminal, I could then just use Claude and now I'm actually using Claude in this terminal.

[6:02]So I say, hey, now this is Claude code. This here is DeepSeek. How cool's that? So let's call this one over to DeepSeek. And of course, all we want to just make sure we're doing here. Could you just confirm that we're using DeepSeek V4, please?

[6:14]Wonderful. So it looks like we were using the flash model, and when I said just use basically this version, it did it. This is how cool this is. It will just update that for us over here.

[6:23]And by the way, if you're brand new to Claude code, I should also call out, by the way, that I've got a full course in the community coming that covers everything from foundation setup, building websites, power features, memory systems, ClaudeBot apps, build anything, design systems, how to monetize it.

[6:36]It's releasing this month, so if you want to go deep, I just want to call that out right now because you might find that helpful in building cool stuff like this. So, we've got DeepSeek now.

[6:43]Now, the interesting thing here is one of the biggest abilities here is its actual ability here to use skills. So let's pull out to a test very quickly. So I want to use my bitly skill, which basically shorten a URL.

[6:54]So let's come down here and actually try that. So I'll say, hey, I would like you to shorten this URL using my bitly skill. All right. And let's come down here and do glaido.com. All right.

[7:02]Which is actually what I'm using to speak type right now, which is really cool. So let's see if it works perfectly. Cause always been my biggest complaint with these things is that they don't actually use skills very well.

[7:11]And if you're using Claude code, it's only really as good as its ability to effectively use skills.

[7:16]I'm yapping. This is too busy just building. It's just building in the background. So let's open us up and see what it does. Cool. Yeah, we open that, that's fine. And just like that, now we're on here.

[7:26]Stop talk stop talking, start typing, if I do say so myself. Freaking awesome. So we've validated it works and we've also validated that you can now use your skills.

[7:33]And in addition to that, we can also use Claude code exactly at the same time. Beautiful. So let's put DeepSeek to the task.

[7:39]But to do that, you need to understand when to use it and when not to use it. So, think about it like this. And this is great, just generally to know this, okay.

[7:47]If you're building anything creative, right now, the best model on the planet for that is Claude. It is, if you like, the design king, we put a king hat on him.

[7:54]If you're doing automations, quick scripts, anything algorithmic problems, unit tests, anything that's kind of like dirty work behind the scenes, that isn't visual or design related.

[8:05]We can bring in the giant whale for that, bringing in Mr. DeepSeek himself. Obviously UI UX polish, we're going to use Claude, multi-file refractors, documentation, blah, blah, blah. We can use Claude.

[8:14]And then code review and audits, we can also tag in Codex. So, to do this, what we're going to do is build a website together. And we're going to do the initial design on Claude.

[8:21]And then we're going to use DeepSeek to do a lot of the heavy lifting in the background. So to do that, I want you to do two very quick things.

[8:26]First one is head over to this GitHub here, which is awesome design md. This is really cool because effectively what this does is grabs the design identity from a gazillion, and that is the correct term for this.

[8:39]Different um like websites and stuff. So this is BMW's for instance. So these are all the design systems of BMW's. If you want a BMW looking website that looks like BMW, you would come onto this website.

[8:48]If you want Apple, then you would come onto Apple. And so and so on, you get the complete idea. So first step is come over to this GitHub here. We're going to click on code and then go straight back over to antigravity.

[8:59]And then we come back over to anti-gravity and click on this Claude here. And effectively, what I'm going to do here and actually since it's a GitHub install, I can just use DeepSeek for this.

[9:07]I'm like, hey there, dude, I'd like to go ahead and install this GitHub repo here, please. All right. Come over and then just paste that one in there.

[9:13]Now, there are some questions about Jack, is this going to China? What about the censorship, Jack? What about Tiananmen Square? What if I want to talk about Tiananmen Square?

[9:22]Well, if you want to talk about that, um don't use DeepSeek, but otherwise, yeah, it's fine. I mean like obviously all these models, like censor in some shape or form, so do be aware of that.

[9:33]One legit call out, though, is that if you do not want to paste naked URLs or API keys into DeepSeek, that is a big, big, big, no, no.

[9:42]So never do that. Don't post this chat here is like sacred. So don't paste anything in here that you wouldn't be okay with popping up on internet.

[9:50]Um that's a general good rule to do with any model, but DeepSeek especially, just don't do that. So this has come back and said do you want to do it? I'll say, yes, clone it, please.

[9:58]Which is cool. And get it to clone that repo. So that's going to be fantastic. Cool. So almost all complete. And again, that was quick.

[10:06]This is the cool thing about using it directly in DeepSeek is I'm not having to wait six business days for this thing to arrive. Like it's just a strap away. So now we've validated and got that done.

[10:10]Let's come over to Claude here and give it a prompt. I'm going to say, hey there, using the design repo awesome designs on the left hand side.

[10:16]I want you to use the Apple design framework to build me a beautiful website about coffee, stay and I also want you to do is lean on and leverage certain design components from this GitHub repo right here. All right.

[10:46]And I'd love to have a calculator that shows people how much productivity they get from drinking coffee. Have that as like the second thing on the website, please. Cool.

[10:54]So we're going to give it a general idea and I'm going to send that off. So okay, Claude's doing the design work, but when it comes to all the other heavy lifting, we're just going to tag in DeepSeek.

[11:03]Now, you could if you wanted to, just get DeepSeek to do this itself. And in fact, I could even show you that, but what I did it, what I basically found is that it just lacked creative imagination.

[11:14]Which means we're going to keep it away from anything visual, tag in Claude, anything that's not visual, we bring in DeepSeek. And once you've built the app, you'll find that so much of it is actually just the backend stuff anyway.

[11:23]And another little hack with websites, if you find a website that just has a design that you think is gorgeous, that you want to kind of lean on to your designs, you can copy that URL.

[11:32]Head over to basically just Google HTML website extractor. You head over to this website right here. You literally drop it in like so, click on view source.

[11:40]Then you'll get the source code for the website, you literally click on download and then save it to your laptop. Then you can pretty much feed that to the model and lean on some HTML for inspiration.

[11:48]So I've built a couple of options for this. This is one very kind of Apple-esque style, which is pretty cool. Come down, you can see from a single batch of your kitchen, again, one idea of a website we could use, if we like the look of it.

[11:58]Another one is what I call Homeland Coffee. So it's just got this cool thing that generates with AI. And again, you come down. This one's really fresh. Get the idea using some good inspiration HTML.

[12:08]This is quite crisp and quite nice. It's got a really beautiful feel and vibe to it. So, Homeland Coffee, something that the boys can definitely trust.

[12:17]So, what we're going to do now that we've got a basic design is get DeepSeek to basically supercharge it. So to do that, we're going to come over to anti-gravity.

[12:24]And what we're going to do very quickly at the bottom of this is we're just going to clear the window like so. Get rid of all the context. And then just give it a prompt. So I'm going to say, hey there, what I'd like to do on this website is add a sophisticated ROI calculator.

[12:37]That basically drives multiple variables. One is going to be amount of caffeine it's consumed and productivity. Make it as sophisticated and as intelligent as humanly possible. Really cool.

[12:47]And then I'm just going to grab the URL for this, so I know what I'm talking about. Come back over and then drop that one off right there. And let DeepSeek do its beautiful work in the background.

[12:56]Now guys, check this out now. I'll be honest, this is surprising to me. It's actually managed, if you look at this, to capture the exact HTML style.

[13:05]This is the cool thing. DeepSeek by itself, if I asked it to do this, would struggle. It'd probably be in crowds in the corner thinking about things. But because I've got an existing brand guidelines and style, it's just replicated.

[13:16]So let's see. I've got daily cups. Let's have a look. And look at this. How where am I at on this? Where's Jack Robert at? Jack probably sits here.

[13:24]Again, five cups of coffee. I want to have the whole bean. Awesome. My hourly rate is, let's just put it all up here to 250. It's a bit above that, but hours worked a day, and let's put that like I don't know, 13.

[13:33]Caffeine sensitivity, I think I'm probably quite low. How many hours do I sleep? I sleep like 8 to 9 hours. And then you get you can see this. Annual ROI is 80 freaking, guys, where do I freaking sign?

[13:44]So you get the idea here. This is all done by DeepSeek at a fraction of the rate. And it's just one example that we could use within DeepSeek.

[13:51]And obviously, if you want the source of how to build websites like this and actually publish them online and host them, I'll put my website master class, so you can check that one out when you get your own home and the coffee.

[14:00]Now, obviously, if you want to use a free model, you can come over to openrouter.ai, create a brand new API key here, for example.

[14:07]Then once you've got that, come back over to anti-gravity and you just say, hey, instead of using DeepSeek V4, I want you to default to a series of free models within OpenRouter.

[14:14]Use the ones that try not to rate limit me and cycle through any of the free models, please. Here is my API key. And effectively, once you restart and follow the instructions, we've now got all of these free models.

[14:24]Quentry coder, Nvidia, Metron 3, Super, and it would cycle through. Now you'd be doing the exact same thing that we're doing with DeepSeek, but instead of DeepSeek, we're using a free model or any other model that we want to.

[14:34]Now, the other thing that I would love you to do, if you haven't integrated chat GPT 5.5 and Gemini into your cloud system, definitely check out this video on screen where I show you the full system.

[14:45]But where this goes from good to out of this world is what we can do now is the following, which is this. Hey there, what I'd like you to do is to tag in GPT 5.5 Codex to review the code to improve it in any way and just make sure that it's actually working perfectly.

[14:58]Now what's cool about this is this leverages your existing chat GPT plus subscription. And you can use dozens of these different things.

[15:07]In that video, I'll show exactly how you set it up. It's super straightforward. And we can tag in the best model, which means we get free reviews, essentially from chat GPT, which in many respects, is kicking butt right now as a model.

[15:18]This will now bring in Codex. Again, I'm doing it in the Claude chat here, but it worked just as easily in DeepSeek. We can bring in any model that we want to. So, with that in mind, now that's the full pipeline, right? Claude for design, DeepSeek for the heavy lifting and then Codex for the review.

[15:31]But where does it really fall down as a model and what's important to know? Well, first of all, your prompts go to China, so again, no Tinaman Square questions.

[15:39]Um again, bear in mind about naked text. Second of all, is UI taste. It is not great at UI. The one caveat I found is if it's already within an existing UI framework, like you saw, I thought that did really well personally, so I think it's good for adding additional things within frameworks.

[15:53]But don't use it as your first design artist. It's no kind of uh Michelangelo if you like.

[15:59]And uh also the the rate is very, very cheap, but there's a promo now. So instead of getting 100 times cheaper, it might be slightly less than that. So, what should I be doing ever right now?

[16:07]Well, basically, if you are corporate, do not use this, because again, it's a corporate database. If you're doing a lot of health care finance, so basically really highly regulated stuff, just keep DeepSeek out of the picture a little bit.

[16:17]You wouldn't want to put in there your, you know, burning, bulging IP that's changing the planet, or like all collateral health care data. But look, if you're a solo dev, using cloud credits and all the projects, if you're learning cloud code workflows, if you're building open source stuff or prototyping, hubway throw scripts, this is freaking awesome for reducing down your costs.

[16:37]Claude for the hard stuff, and then tag in obviously chat GPT to do the reviews. Now, DeepSeek is reducing our cost, but we now need to increase performance, which we can do by using super skills by watching this video right here.