Graphify DOUBLED My Claude Code Token Bill (4 Real Fixes)

Last week I discovered a viral AI tool that claims it can slash your API cost by 70 times. Now with how expensive Fable 5 is, this is absolutely a game changer. And over on GitHub it has 98,000 stars. So of course, naturally I had to check it out for myself. But here's the kicker, the results may surprise you. So follow along for this tutorial here. All right, so the tool is called Graphify and here is their GitHub repository. And how it works is that it will make a map of your entire code base first. And so all of the functions that are related are in one segment of this code graph. So everything that is in one color is all of the related functions. And so the whole idea here is instead of Claude having to read your entire code base, it only reads the sections it needs to. And so it doesn't need a lot of computation power because it's not reading your entire code base anymore. It's only reading the specific sections that it needs, okay? And here is another map of how it works. Basically what Graphify does is that it will make a graph of your entire code base first, and this is just a one-time thing that it will do. And from then on, you will save on token costs because the entire graph of your code base has been made so that Claude doesn't have to read through your entire code base every time you put in a new prompt. Okay, so the premise of it sounds really interesting, and the graphs are obviously visually appealing. So this seems like a great idea, right? Alright, so I put it to the test. Here's what I did. Alright, so in order to put Graphify to the test, here is what I'm going to do. So first, I'm going to download this code repository. So it's cal.com. Essentially, when I download this code repository, I get an open source version of a calendar app. And this is what it looks like. And the reason I'm doing this is, first, I'm going to get the base of this code, and then I'm going to put in a prompt to modify this codebase, and I want to see how many tokens it took to make that modification. And then what I will do is install Graphify, and then I will execute the exact same prompt, and I want to see how many tokens I was able to save and how much money I was able to save. And this will help me determine if Graphify will actually produce the 70 times in API token costs that they claim, whether it's even close to that. Okay, okay, so I am actually in VS Code right now, and I linked it up to Claude, and you can absolutely do this in terminal as well. And basically what I'm doing is just the first step where I am cloning the calendar repository. Okay, and I'll be right back after the calendar GitHub repository is cloned onto my local machine. Okay, so after I cloned that repository over in VS Code, I actually switched over to terminal here. And the reason that I did that is so that I can see how many tokens I have been using. I can't do that inside of Claude Code or VS Code. So that's why I switched over to terminal. And in retrospect, I should have just started with terminal to begin with. But you live and you learn. All right. So basically, I got into the calendar folder, and I opened up Claude in terminal so far. So that's what you're seeing here. Then I put in this prompt right here, and I will put in the exact same prompt after I install Graphify as well. This prompt is, I want to reduce no-shows. Trace how meeting reminder emails work in this codebase, what triggers them, when they're scheduled, and every file involved. Then tell me exactly what I need to change to send a second reminder 1 hour before the meeting. You can see Claude going to work now, and it's going to ask me for a couple of approvals before this entire prompt finishes processing. And then after this, we will then go ahead and install GraphiQL and run this exact same prompt again and see if there's any token savings. Okay. All right. So I'll be back in a minute after this finishes. Okay, so it's been about 5 minutes-ish or so by now. And Claude has finished running our prompt. And so what we're going to do right now is exit Claude Code in terminal. So the way that we're going to do that is we're just going to type slash exit, and we're going to go back to the main terminal to see how many tokens that we have spent so far. Okay, so I'm going to type in slash exit, and I'm going to put in this command right here: npx space cc usage at last session, and then since, you know, the date that I'm executing this on right here. I can include all of the prompts and instructions I've used below, and you can see right now it has calculated how many tokens I have used for this session. So I have used Claude for some other sessions as well. So in order to know how many tokens I have used for this session, I put in this command line that tells me what the last session was. Okay, so this gives us our baseline. And I'm using Fable-5 for all of the models here. So if you do want to switch models and see which model you're using, just use slash backslash model, and you can choose or update the model that you are using. So for this particular session, you can see we have used about just under 600,000 tokens, 587,569 tokens to be specific, and it costed us $2.25. Okay, so that is our baseline. This is the amount that Graphi has to be able to beat in order to achieve any kind of cost savings. All right, so let's go ahead now and install Graphi. So to install Graphi, this is the command line that I put in: uv tool install graphi. And the next line is graphi install. Of course, I will include all the instructions in the resource section below if you do want to give this a try. And then this whole thing is going to take a while to install the entire package, so I will be right back. until this entire package is installed. All right, it's been a hot minute and Graphi has finished installing. And what we're going to do now is start a new Claude chat window inside the terminal and we're going to put in our exact prompt inside. And these are the command lines that I put in to basically start a new Claude window. What I'm doing here is basically checking, does the Claude skills include Graphi now? Because I want to make sure that Claude knows that it now needs to use Graphify in order to process code because it's going to create a whole map and everything else, right? So the first thing we're going to do before we run the prompt though is that we have to create the whole map, right? That's the whole premise of using this tool. You got to create the map first. So to create that, we first call Graphify. So backslash Graphify, which will call this skill. So you can see that the word Graphify turns blue. So that again tells us that we have now installed this properly and it's just space, and then a dot. Okay, and so this is what triggers GraphiQL to go through our whole codebase, in this case is the cal.com repository, that's the one that we just downloaded, and it will scan the whole code repository and create a map around that. Okay, and we only need to do this the first time, and after that it will know to use the GraphiQL map, and the whole premise is that it will start to save us API cost. We'll see if that actually happens or not here. Okay, so let's wait a couple minutes for Graphite to create our map, and I'll be right back once that is done. Okay, so after the graph has been built out in the background, so what I'm going to do now is put in the command line to have the graph being displayed to us. And this is the line on the command line that you have to type in in order for your graph to show up. Again, I'll include this in the resource section below. And again, we're going to wait a couple more minutes for this to finish creating. So I spent roughly 10 minutes or so, so far, waiting for the graph to finish computing. Okay. And also right here, I am also installing the Graphify skill into Claude as well so that it knows how to display the Graphify graphs here. All right. So we're going to have to wait a couple more minutes for this to finish computing. Okay. And here is the final command line that I put in to display our Graphify graph of our calendar app. And then I open it up. And this is what it showed up as on our browser. And you can see the different codes being mapped in their respective segment. And we can see what the color codes represented it. And I have also created a different view. So it actually gave me like 2 different views because I tried this 2 different times. One was a different color scheme. And this one was a different color scheme from that. But it was the exact same information being fed into it both times. So our graph worked, which is awesome. Okay, so what I also did was open up the graph report. So this is like a text version of what we saw on the graph right there. So that also worked. So all of the confirmations are saying that we now have GraphiQL installed. Okay, so now I'm going to exit. And what I'm going to do is see how many tokens it cost to install GraphiQL here. So using the same command line, Here, the NPX one. Now I want to see how many tokens this particular session cost us. And then lastly, I want to compute how many tokens once we install Graphify, the exact same prompt will cost us in the future, right? And as you can see here, so I used Fable 5 to install Graphify, which in retrospect was probably not a good idea. I probably should have used like Opus 5 to do that because it wasn't like something that required a lot of thought for the AI to just install Graphify. But anyway, so you can see this cost over 10 million tokens in order to compute and $28, right? So I could have easily cut that in half though if I just switched models and switched it over to OpenSwipe. But you know what, you live and learn, okay? So that was pretty expensive. However, this was a one-time cost, so if it does actually produce the amount of savings that it promised to, then it could be worth it. Okay, so now I started a brand new Cloud Code now that we have Graphify installed. And I'm going to enter in the exact same prompt. Now, let's run it. And then let's exit Cloud Code and see how many tokens this particular session cost. Okay, so that is the exact same prompt as before. And we're going to wait a minute for it to finish loading here. All right, so it's been about 4 or 5 minutes or so, and we have finished running that prompt. So I'm back in Claude now, and I put in the same npx command, and it's going to pull up the Claude sessions as well as how many tokens have been used. And here is what surprised me. So this session was the exact same prompt As you can see here, this was— you can see here the last session that we've used is the one starting in F1C59, etc. And this particular session actually cost $5.61. So it's more than double the original, which was $2.25. So we spent $28 installing Graphify. And the exact same prompt was double in cost, not less in cost. And at this point, I was scratching my head and I was kind of like going like, what the heck happened? So I'm going back and forth with Claude and I fed it the entire terminal transcript and I asked it, okay, what the heck did I do wrong and what am I missing? And Claude told me, well, maybe you forgot to call Graphify. So perhaps Claude, when I was computing this, it didn't actually use Graphify. So it bypassed the graph it created and just computed the prompt normally. And perhaps it took a different route. So that's why it was more expensive. So I went, okay, well, that seems like a reasonable explanation. So what I did was I created another Claude chat. And this time, I made sure that I actually called Graphify. So I started off this prompt by doing the backslash Graphify. And you can see that the word Graphify actually turned blue. Now I'm specifically telling it you need to use the Graphify skill that I just installed into you to process this prompt. It's the exact same prompt as before, and we're going to let this run, and then we're going to exit the Claude chat again, so backslash exit, and then we're going to check our token usage again using the exact same npx command. So we'll be right back once that happens. All right, Claude has finished computing our prompt request now using GraphiQL. So let's exit out and let's go back to the main terminal and see how many tokens we used this time. Okay, so I am just copying the same command line before. And what this ls command line tells me is it tells me what the session code of the latest session was, which is what I need to know in order to determine which of these lines on this graph here represents the latest session, right? And it is the one starting with 85A8C. So the 4th one down on this chart. And you can see here, it actually cost $3.95 in order to compute the exact same prompt. And execute on that exact same prompt. So at this point, I was really confused and I didn't even really know if I wanted to make this video or not, because while I saw a lot of like stuff online about how great this tool was, when I actually tried it out both times, the cost was way higher than just using Claude Code plainly, right? As you can see, the result before using Claude Code was $2.25. It cost me $28 to install Graphify, and after running the exact same prompt twice, the first time it was $5.61 and the second time was $3.95. I have no idea what I'm doing wrong. So if you are a technical person and you can see what my methodology was wrong, please let me know in the comments below. I would love to know. But from my experimentation, Graphify actually cost a lot more. Like, it's a cool concept, But it doesn't actually work. Like, I think it's just like a cool visual graph and people go like, oh wow, that's a cool visual graph. But like, from all my testing that I have done, which I agree it's not the most comprehensive testing you can do, but I'm not seeing any cost savings. I actually see my token cost like exploding, like doubling from using this tool versus not using this tool. So I'm not seeing what all the hype is about. Okay, so that's just me. If you can see what I'm doing wrong, let me know in the comments and I'm happy to make a follow-up video to this if needed. Okay. All right. Before I wrap up this video, though, I do want to give you some tried and true techniques to actually lower your API costs. So the first one is model routing. So depending on the difficulty of the task that you're trying to do, choose the right model for the specific task. I mean, you saw my mistake from earlier, right? I used Fablefy to install Graphify, where I could have used like a cheaper model like Opus-5, which is half the price, to install GraphiQL. That task I didn't need the most powerful model to do. So by choosing models that are less powerful for tasks that doesn't need as much computation, you'll be able to save on API cost. The next way that you can save on API cost is by using the /clear and /compact regularly as you're coding. In Claude code. So what this does is that it compacts your previous conversations or clears the previous conversations. Yes, this means that Claude will not have as much context as what you have done before. But by doing this regularly, you'll be able to save on API costs. Because if you don't do this, what Claude does is that it reviews everything that you have just talked about before, which is a lot of API tokens. And then it looks at your new prompt, and then it computes its response. So if we have talked about previously, it's not relevant to your current prompt, then use /clear and /compact in order to save on some API usage. All right. The 3rd technique that you can use to lower your API costs is to have a very lean Claude.md file. The Claude.md file will include all of the instructions that Claude will go through for every single prompt. So you only want the most important information that Claude needs, like the context it has to have for the particular project at hand. So if it's not something that Claude needs for every single prompt that you're going to run, then don't put it in the Claude.md file. If you did not start a Claude.md file, type in /init, so I-N-I-T, which stands for initiation, in order for Claude to create one for you automatically. And then you can type in /memory to see what is in your Claude.md file. If you're working on the local desktop version of Claude Code, you can navigate to the folder that you are working in, and it should be in the root folder, the cloud.md file. So take a look at the prompt in there and make sure it's lean and that it's actually the information that Claude needs to run every single one of your commands. Because the more instructions you give it, the more API, the more tokens it's going to take to look over all of your instructions, plus the prompt at hand and then compute everything, okay? So regularly review your Claude.md file and make sure it is lean, solid, and filled with information that Claude needs to know. Now, the last technique I'm going to talk about today is using a deterministic workflow instead of using AI for everything. So here's what I mean by that. A lot of people get automation and AI confused, and they think that because AI can do everything, they should use AI to do everything. And this is the whole concept that emerged called token maxing, the idea that you will use AI for everything, right? But here's the question that you have to ask yourself. Did this step need AI at all? Let me give you 2 examples to really illustrate this, right? So let's suppose you have a new lead that came in, and then after that lead came in through your CRM, you want to send them an email and you want to send them a text message. If you use an AI agent to do that, the agent is thinking, okay, what tools do I have in my tool belt? To send an email. Okay, great. I have Gmail. Okay, let me send that email. Right. But this is going to cost you a lot more tokens than using something like Zapier and just creating an automation. Lead comes in, the next node or module is going to send an email. After that, it's going to be send a text message. Right. So the question you have to ask yourself is that, is this a workflow that is an if-else? If this happens, then do this. Right. So if it is the same steps that happen every single time, what will allow you to save on tokens is actually making an automation instead of using AI for everything. Now, going back to the same example here, so where you can plug in AI and where it makes sense is let's say that lead comes in and you want to generate a unique email, customized email for every single lead that comes in. And so you can use AI to generate that custom email through a ChatGPT or Claude module in Zapier, and then you will still use a deterministic workflow to send it out. So what allows you to save on tokens is by having an automation for the deterministic, the workflow part that doesn't require any thinking from the AI, right? So by combining automations with AI in the appropriate order, it will actually allow you to save money. In the long run. And this is where the whole concept of value maxing instead of token maxing comes from. Let me give you another example to really solidify your understanding here, right? So let's suppose you get an email from one of your members, and it has the word upgrade. And whenever someone wants to upgrade the package from you, you will then assign that particular email ticket to, let's say, your enrollment specialist. Let's suppose her name is Laura. Okay. And then, and so because that specific path happens every single time, if email contains word upgrade, assign ticket to Laura, that is a deterministic workflow. Okay. If this, then that. So like there's nothing to decide. You just want a specific action to happen if the condition is true. Okay. So in this case, it makes a lot more sense to use a deterministic workflow using an automation tool like Zapier to do it instead of having an AI agent to do it. Now, where it makes sense to have an AI agent in this example is that the AI agent can actually draft up a new quote or pricing for that member and put it in draft mode. And then Laura can review that email with the upgrade proposal for that member before before it actually goes out. Okay, so again, value maxing is the idea that we are trying to minimize our total usage cost, and we're not trying to token max, which is to use AI for everything. All right, hopefully that concept is now clear. By using these 4 different techniques I just discussed here, you will actually be able to lower your token cost. I hope you enjoyed this video, and unfortunately, I was not able to save on any tokens with my Graphify and I have no idea what I did wrong. And also, if you like this video, please like and subscribe. That will help out the YouTube algorithm so much. And I also have a free course that will teach you how to build these AI agents in order to increase your productivity, save you on cost, and increase on your profitability as well. Make sure you sign up and check it out below. And again, I'll include all of the resources we talked about today here as well. Lastly, if you are a business owner and you want to build out your entire AI workforce and AI-ify your entire business, Book a call with me below, and I can't wait to see you very soon. Thank you guys so much, and I'll see you next time.