Why I Cancelled My Claude Code Subscription

I was a CloudMax subscriber. I paid $200 a month. I loved it. I was a beast power user using hundreds of millions of tokens per day. Sometimes I can easily say that just over the last 12 months, CloudCode went from pretty neat to OMG can't live without it. Something as a builder and solopreneur, I genuinely couldn't do without. And that is exactly why I canceled my subscription, because I realized there was something seriously wrong with the dependency on Anthropic. Let me explain. You see over the last 30 days of my subscription, on two of them, I surpassed 700 million tokens per day and other days were well on their way. If we do all the math here with the general API costs, those two days alone would have cost over $50,000 around $2,000 if you factor in all the token caching, but we'll get to that later. But even with the discounted cash tokens, we're still looking at approximately $1,800 just for two days of usage. And my subscription only costs $200. Something is seriously wrong with the economics here. And I don't want to be hooked on Claude when the other shoe drops. How was I getting all of this for $200 a month? Someone asked to foot the energy bill here. AI is not free yet. The more I use Claude code, the more my work in business depended on it. Until I finally started thinking about what would happen if that $200 plan disappeared next week. What if they changed the limits or the product itself at that point, how easy would it be for me to leave? And over time, that question started bothering me a lot because a large part of my life was revolving around Claude. Fortunately, there's a way out of this today. We have open rate models that are not only surprisingly good, but they offer results comparable to Opus and Fable. Meanwhile, there are more and more open source agent frameworks and harnesses that are model agnostic, which allow you to get creative past the handful of models offered by Claude or open AI or any American frontier family of models. So I decided to try something. I want to see whether I can replace my $200 Claude subscription with open models and a smarter workflow without giving up the things that made Claude so damn useful to me in the first place. Let's start with the token inference. As I said, those two days alone would have cost approximately $1,800, but that's only two days of my 30 day subscription that I pay in total $200. So I started digging into that because it seemed strange. The API prices are public. If you took the amount of input and output tokens from those days and price them as normal API usage, well, you see these high numbers in the tens of thousands. So either their API margins are obscene or they're burning VC money to subsidize user inference, get us hooked and hiding the actual cost of delivery here. It's probably a little bit of both. They have a good product. People want to pay for it. Claude code has become one of the most useful products some of us have ever paid for in our lives. We use it constantly. Someone like me, I could sit down on a weekend, start building something and not stop for 20 hours straight. The more you use it, the more you got from that easy $200. That's an incredible deal for a heavy user, but there is a flip side. Once you get used to working that way, you build your entire workflow around it and you become dependent. You stop thinking about the cost of every request. You stop worrying about which model you're using. You just open cloud code and keep working. And I had definitely gotten to that point lazy with my AI usage because the cloud code harness itself with the subsidized inference allowed me to just work and not have to really think about anything under the hood. So when I started looking at how much of my actual work was being done with Claude, I started wondering what happens if the deal changes because we've seen AI products or anything in tech change their limits before. A plan that looks generous today can look very different once the economic reality of running a business kicks in, which is often what happens when a company is about to IPO. But that's another story. So I'd already seen how quickly my usage could explode with over 700 million tokens in a single day. That's just what my workflow started to look like. So I had to ask myself the uncomfortable question. If Anthropic changed the price tomorrow or tighten the limits or change the way cloud code worked, how much of my workflow would I have to rebuild or start from scratch? And that question bothered me a lot more than the $200. So I used cloud code to do a lot, build landing pages, sales pages, automated meta ad campaigns, generate visual assets for videos or presentations with remotion and hyper frames and handle parts of my content workflow, including repurposing metadata generation, et cetera. I did a lot. And then there were my side projects. I had a project I was working on with my son where he created a world, built a story and we turned it into a book and a YouTube channel and an educational alert to read game on its own website. We made our own vibe coded Patreon and cloud was sitting right in the middle of that. If I didn't have the subsidized free inference essentially to build that I wouldn't because it wasn't guaranteed to make any money. It was just a fun project. And it was with this project with my son. I realized I would open cloud code before I'd even think about whether there was another way to get the job done. I already knew how to use it. The tools were there. The context and memory was there. I built up all my workflows around it. So starting something new was easy. And that convenience adds up as an ignorance debt that eventually we must pay back because you've stopped questioning the tool because it just works. And after enough time, you inevitably feel locked in whether you realize it or not. The cost of switching just becomes too high. And that's what Anthropic wants. That's what any business wants. It's good business because it turned into even the thought of me canceling my cloud subscription was more than just cutting another subscription from my life. I'm not just going to stop building with AI. I have to figure out an alternative. There's a cost to switching. Some of my workflows would be easy to move, but some of it would not. And truthfully, I didn't like how little I'd actually thought about that before I'd spent all this time making cloud more useful to me while quietly making my own workflow harder to separate from it. And that's a weird position to end up in. You start with a tool because it's convenient. Then you build around it. And eventually that convenience becomes the reason you don't want to leave. It's like the relationship, you know, you need to end, but it's just too difficult. So eventually two things were really bothering me. The future economics of the subscription and the company, the vendor lock in the fact that I've built so much around a system that I didn't really control. So at this point I asked myself, I don't want to just be stuck with Claude and I don't want to just resign to that, especially with all the geopolitical drama Anthropic has been cooking up. So I started looking for a way out. Now the question arises, what the hell do we use instead of Claude or open AI? Cause I'm talking about Claude here, but the truth is this applies to any AI provider, especially one that doesn't respect your privacy and is giving you something for free. Remember what they say? There is no such thing as a free lunch. And if lunches were tokens, well, I've gotten a lot of free lunches with Claude code. But the answer to this question is not to replace Claude or Codex with one thing. And that's how I was thinking about it the wrong way. At the beginning, I was looking for a single tool that could replace Claude. Oh, is Codex better? Is Gemini CLI better? But that doesn't really make sense because why should one model or one family models have to provide everything? If I'm writing code, maybe I want just one model. If I'm planning a project, I might want another. And this is where things got interesting because the options had changed a lot since I first started using Claude code. Open weight models now can do some serious work from Kimi K3 to GLM 5.2, GLM 5.3, DeepSeek V4, Quen 3.8. And you can connect all these to open source agent harnesses and a tool like open code, which has quickly become my cloud code replacement is what they call model agnostic. I can connect any provider, any model into there, and it's completely compatible with my cloud code agentic workflows. So once I started thinking about it that way, the whole setup changed. I was able to start using Kimi K3 as more of a planner and orchestrator. It's more expensive, but it's super intelligent and it's actually a lot friendlier and easier to work with than fable. Meanwhile, GLM 5.2 is cheaper could handle a lot of the coding and it's a very good creative writer as well. And for even simpler jobs, DeepSeek V4 flash may as well be free. It's fast, intelligent. And while I wouldn't want it planning my projects, it can handle the menial tasks perfectly. And getting comfortable with this meant that I could move away from the belief that I need to cloud code to achieve the productivity that I desired. But there's another advantage I want to double click on, and that is actually choosing where the models come from. Now, as I was talking about when I'm creating something for my business content, like what does the privacy really matter? Well, in a lot of ways, the output itself, whatever privacy, no privacy, whatever. But there's the actual creative process that comes from me, my unique way of thinking as a human being that is being profiled by tools like cloud code codex Gemini. They're learning my psychology and making a profile about me. And once you start going deep into that, it's kind of freaky. So there's another reason to shift from these proprietary frontier services. And that begins to align with the inevitable next step. If you're a builder with AI is what you can start using these harnesses for inside your own house. I started messing around with home security systems with raspberry pies in the living room, hosting video game emulators. And once I hit the stage, I realized I don't want anthropic in my home network. No, thank you. And the models I can host locally just aren't there yet. So that's why I use Venice as a provider of all of my favorite open weight models because they have zero data retention agreements with the providers, meaning nothing is being stored by the provider of my inference. No profile is being built. And even if there's a nefarious entity inside the GPUs that my provider doesn't even know about, they're not going to know what's coming from me. So this started creating a comprehensive view of how I want to approach my AI setups. Now, instead of opening cloud code and asking you to do every little thing I can now build workflows that use different models, different tools, different providers for different parts of the job. And that brings us into a topic that most cloud code users have probably not thought about much. And that is smart model routing. I had no choice but to start digging into this. I'd gotten into the habit of treating cloud like the default answer to every problem. So once I stopped doing that, the economics actually started to look very different. I could have one model think through what actually needs to happen and then hand off the plan to another model for implementation, a cheaper model. And it's not a coding thing. I could use even simpler models. And then when I need a serious final review or audit, then I bring the big boys back in. So the most interesting part of something like smart model routing is that it can be handled in the workspace itself. So what might not be obvious as a cloud code user is that your AI agent can make calls to other models with instructions, with a system prompt, whatever. And what's really interesting about this, if you've only used cloud code or codex and you've never built your own workspaces, is that the workspace itself can handle a lot of this smart model routing. The better your instructions are and agents MD, the smarter your agent can be in sending certain tasks to certain models. This will not only give you better performance and results, but it will save money, which is important when we're no longer using the subsidized inference by anthropic. A well-structured workspace means my agent is going to be smart about the tasks it does versus the tasks it sends sub-agents to do. So as an example, I could use cloud code for a task. I could use open code for a task and they might use a completely different amount of tokens, even if I'm using the same model. So let me explain. If I throw everything at cloud code and say, Hey, figure this thing out, just do it. It'll probably use a lot of tokens in a fresh workspace with no instructions. It'll make its own instructions. It's so smart. You'll figure it out. It'll probably do a good job, but it's going to burn through way more tokens. Then if you use the same model, Opus Fable in a tight workspace with instructions, with skills, with sub-agent profiles, et cetera, because it has context to work with not only that, but within the model workspace, you can integrate model routing. So if something needs to be done by a different agent in your instructions, you simply say, when we do this thing, send it to that model from that provider. Here's the API key done. You could do this with cloud code, but why would you even think to do that when cloud code is just so good and you're not paying for the actual inference and a whole generation of AI users are going to be handicapped when reality comes knocking. And I don't wish that for you. So I'm glad you're watching this. And now is probably a good time to subscribe to my channel. If you do want to learn how to make workspaces that are not dependent on cloud code, because that's what I explore in my videos on the channel. So thanks for hitting subscribe. So once I started understanding the different economics of running different models in a workspace, I wanted to know how far I could take it. Could I actually rebuild enough of my workflows that I wouldn't miss cloud? The answer was way more obvious than I could have expected. So I decided to actually test it out. I've canceled my cloud subscription. It's time to move on. I'm running a Venice max account, which is also $200 a month. And it gives me $225 worth of credits with their service. But there's over 300 different models that I can choose from from the top of the line, open weight models, like Kimmy K3 to GLM 5.2 to deep seek V4 flash, et cetera. There's a lot to play with here that will allow me to save money and retain performance. It just means sacrificing the convenience. My goal is not to recreate cloud perfectly, but it is to get similar results without having to spend a lot more money. The whole point of my experiment is to find out where the gaps in my own workspaces and workflows are. I have to be present with every single word and every single prompt because all of those are costing me tokens, which costs money. If a task doesn't need the strongest model, I shouldn't send the task there. If a bunch of context is going to be used, can it be cached? If I'm recycling instructions, I need to make skills and put them in my behavioral context files. This is the stuff I'm most interested in because this is the true architecture of owning your own AI. Sure. I might not be running the actual intelligence on a machine I own, but theoretically in the not so distant future, I could. And more importantly, as long as there are vendors out there willing to provide for me, the models at a cost that I want to use, then the most important aspect for me is owning the workspaces themselves because they can be mine and no one can take those away from me. No one can shut those off. I'll just find another model provider. We live in a world of markets. Find the service provider that works for you. This way, if a particular tool or model or provider stops working for me, the rest of my workflow, the rest of my life with AI doesn't have to come down with it. I don't know yet how far this will go, but that's what the next few weeks of my life are going to all be about. Not using cloud code and moving into agentic workspaces that I control with inference that I'm paying for every single prompt accounted for a responsibility that I have chosen to stop outsourcing to the corporate provider of my artificial intelligence. Can open weight models, open source agents and smart model routing replace what I was getting with cloud code? Or am I going to end up resubscribing the clod tail between my legs in a few weeks from now, that's the experiment stay tuned because I will be posting updates, letting you know how it goes. Finally, if you're interested in learning how to build these agentic workspaces yourself, I invite you to check out the AI captains Academy. There is a whole curriculum inside for understanding from the ground up how AI agents work, how to select between models and providers and agent frameworks, et cetera, how to approach building your agents, including knowledge infrastructure and context engineering. We do two calls a week. One is a workshop. One is a questions and answers session. Come say hi, be in a community of builders that use AI, that love AI like you and me. And we share our solutions to these problems and challenges we face such as replacing cloud code. So if that sounds interesting to you, check out the link below and I hope to see you there. Thanks for watching. And I hope this has inspired you to reconsider your own clod or open AI subscription.