For 12 months I paid $200 a month for Claude Max, and I’d have told you it was the best money my business spent.
I was a beast power user — hundreds of millions of tokens (the units AI models bill you in) on heavy days. Claude Code went from “pretty neat” to “can’t live without it” somewhere inside that year, and I built my whole workflow on top of it.
And that is exactly why I decided to cancel my Claude Code subscription.
If you’ve been weighing the same move, I want to show you the math that decided it for me, the audit that bothered me more than the price, and the setup I’m running instead. Because the price was never the problem.
The math that made me cancel
Over the last 30 days of my subscription, two of those days crossed 700 million tokens each. Other days were well on their way.
Price that usage at normal public API rates and those two days alone would have cost over $50,000. Around $2,000 once you factor in token caching (you pay less when the model has already seen the context). Roughly $1,800 even on the discounted cash tokens.
Against a $200/month subscription.
So either the API margins are obscene, or somebody is burning VC money to subsidize user inference, get us hooked, and hide the actual cost of delivery... Probably both. A plan that generous today can look very different once the real economics of running the business kick in — and companies tend to do that math right around IPO time.
I’m not saying the plan is bad. The full cost picture, including the $20 way to start, is in my post on what Claude Code actually costs. For a power user it was an incredible deal. I sat down on a weekend, started building, and didn’t stop for 20 hours straight, and the more I used it, the more that easy $200 gave back.
That’s the trap. The deal was real — and so was the dependency it built.
The incredible deal, and what it quietly built
Once you get used to working that way, you stop thinking about the cost of every request. You stop thinking about which model you’re using. You open Claude Code and keep working.
Lazy with my AI usage, because the subsidized harness let me work and never think about anything under the hood.
(I know how that sounds. But getting lazy is the point of a good subscription... which is exactly what makes it dangerous.)
The moment I saw it clearly was on a project with my son. He created a world, built a story, and we turned it into a book, a YouTube channel, an educational game on its own website, even our own vibe-coded Patreon. Claude sat right in the middle of all of it. And I caught myself opening Claude Code before I’d even considered whether another way to do the job existed.
That’s when a phrase formed that I can’t shake: convenience adds up as an ignorance debt (everything you stopped questioning because the tool “just works”), and eventually you pay it back.
I’d spent all this time making Claude more useful to me, while quietly making my own workflow harder to separate from it.
The question that bothered me more than $200
Here’s the audit I finally ran: if Anthropic changed the price tomorrow, or tightened the limits, or changed how Claude Code worked, how much of my workflow would I have to rebuild from scratch?
Some of my workflows would move easily. Some would not. A large part of my working life revolved around one vendor, and I hadn’t thought about that even once while I was building it.
And that question bothered me a lot more than the $200.
I spent weeks asking the wrong question
My first instinct was to hunt for a single tool that replaces Claude. Is Codex better? Is Gemini CLI better? I compared specs like I was shopping for a phone.
Wrong frame. Why should one model, or one family of models, have to provide everything? Planning a project might want one model. Writing code, another. Simple manual jobs, an even simpler one. Final review, the big guns again.
The question that actually matters: which workflow replaces the dependency? Once I asked it that way, the whole setup changed.
The exit path I’m running
Three pieces. A harness, the models, and the routing.
The harness. I moved to OpenCode, which is what they call model-agnostic (it connects to any provider, any model), and it’s compatible with the agentic workflows I already had. If you’ve never built your own, that’s the part I teach first, and my week-one harness guide walks the whole build. And this isn’t theoretical for me: I already ran a real client job on the open-source stack for $9. Receipts and rough edges are in my Claude Code alternatives post.
The models. Open-weight models (models whose weights are published, so anyone can run them) got seriously good while nobody was looking. My working split: Kimi K3 as the planner and orchestrator (more expensive, super intelligent), GLM 5.2 for the bulk of the coding (cheaper, and a very good creative writer), and DeepSeek V4 Flash for the manual tasks, where it may as well be free.
The routing. This is the part most Claude Code users have never needed to think about, because they’re not paying per token. Smart model routing means your workspace decides where each task goes. It lives in the files: the better your instructions are in your AGENTS.md, your skills, and your subagent profiles, the smarter your agent gets about which tasks it does itself and which it sends to a cheaper model.
Same model, same task: in a fresh, empty workspace it burns far more tokens figuring things out than it does in a tight workspace with instructions and context to work from. So a well-structured workspace gives you better results and lower spend. That combination is invisible while someone else pays your inference bill. It becomes critical the day you start paying your own.
Then there’s the part I didn’t expect to weigh this much: privacy. Tools like Claude Code and Codex aren’t just answering my prompts — as I see it, they’re profiling my creative process and my psychology along the way. That bothered me most once AI reached my home lab. I run Raspberry Pis for home security and game servers in the living room, and at that point the line drew itself: I don’t want an Anthropic in my home network.
Local models aren’t there yet for my work (honest limitation), so I route through Venice, a provider with zero data retention (nothing is stored, so no profile gets built). Their Max plan is also $200/month, it comes with $225 of credits, and there are 300+ models to route between. Same price, very different deal structure.
What owning the workspaces actually buys
I’m not running the intelligence on hardware I own yet. Someday I could be, and I’m building toward that. But as long as vendors compete to sell me models, the most important thing I own is the workspaces themselves.
They’re mine. No one can take them away. No one can shut them off. If a tool, a model, or a provider stops working for me, I find another one (we live in a world of markets), and the rest of my workflow doesn’t come down with it.
That’s the real reveal of this whole experiment: the asset is the system you own. A system you own survives every model swap and keeps running the business while you sleep on it. That’s how your time becomes yours again.
So... did cancelling work?
I honestly don’t know yet. That’s the experiment.
I’m a few weeks in, every prompt now costs me something, and the goal is to find where the gaps in my workspaces and workflows are. There’s a real chance I’m back on a Claude plan with my tail between my legs before the month is out. (If that happens, you’ll hear it here first.)
What I already know: the question of how much of your workflow you’d have to rebuild if the deal changed tomorrow is worth answering for your own setup, whichever way you land.
Your 15-minute version of my audit
You don’t need to cancel anything to run this check. Give it one focused quarter-hour:
- List what runs through one vendor. Open your main project and write down every workflow that touches it: content, pages, automation, the fun side project. Most people stop listing somewhere around item six.
- Mark the rebuilds. For each item, ask: if the deal changed tomorrow, is this a 10-minute swap or a start-from-scratch?
- Pick one to make portable. One workflow, moved into files you own: instructions, skills, the works. That single move starts your harness, and it compounds from there.
That’s the whole audit. It costs nothing, and it tells you the exact number I didn’t know I was avoiding: how much of your business one company can take offline.
If you want to build agentic workspaces you own (choosing models and providers on purpose, routing by task, intention first), that’s exactly what we work on inside AI Captains Academy. Two live calls a week (one workshop, one Q&A), a crew of builders figuring this out together, and the full Armory of battle-tested Rigs included at every tier ($83/month or $497/year, 7-day free trial).
Cancel the subscription or keep it. Stop building your whole business on something you can’t swap.
Questions people actually ask
- How do I cancel my Claude Code subscription?
- Claude Code comes with Claude Pro and Max plans, so you cancel by turning off the plan in your Anthropic account's billing settings. Download anything you care about first, then cancel. But know that your project files, instructions, and skills live on your own machine, so the important parts leave with you.
- Will I lose my work if I cancel Claude Code?
- No. And this surprises people. The part you own is the harness: your project folders, the instruction files, the skills, the workflows. None of that lives inside the subscription. What you lose is access to the models. Swap in a different model through a different harness and most of the work keeps running.
- Is Claude Max worth $200 a month?
- For a heavy user, honestly, yes. It's an incredible deal. I pushed through two days that crossed 700 million tokens each, work that would cost roughly $1,800 at cash API rates, all on that $200 plan. The real problem was never the price. It's what a deal that good quietly builds around it: a workflow you can't easily leave.
- What should I use instead of Claude Code?
- There's no single replacement, and that's the trap I fell into first. The working pattern is a model-agnostic harness like OpenCode, open-weight models (Kimi K3 to plan, GLM 5.2 to build, DeepSeek V4 Flash for cheap manual tasks), and routing rules written into your workspace so each task lands on the right model.
- What is smart model routing?
- Smart model routing means matching each task to the model that fits it instead of sending everything to one expensive default. One model plans, a cheaper one implements, an even simpler one handles manual jobs, and the big models come back for final review. Written into your workspace instructions, it improves results and lowers spend at the same time.
- Why did Jordan Urbs cancel his Claude Code subscription?
- The math and the lock-in, in that order of discovery. Two days of his usage would have cost about $1,800 at discounted API rates against a $200 subscription, meaning the price was subsidized. Digging deeper, he found his whole workflow was built on one vendor he doesn't control, so he moved to open-weight models and a harness he owns.
