I paid $200 a month to get lazy. And not the visible kind.
I was producing more than I ever had. Landing pages, ad campaigns, media repurposing, on top of a normal 40-hour week, plus 10 to 15 hour weekend sessions.
The laziness was somewhere harder to see. It was about what all that building was actually costing me, and who was paying for the part I couldn’t see.
A few weeks ago I canceled my Claude Code subscription. Not to jump to Codex. To go fully open-source: an open harness (the app that connects you to the model) and open-weight models (models you can download and run yourself). I wanted to find out whether I could still do the work without the $200/month plan.
The math is what started it. There is no way a user paying $200 a month maxes out billions of tokens in 30 days and covers the energy bill. There’s just no way. Which means the inference (running the AI) is being subsidized. The company sells it below cost and eats the difference.
And the rules are already changing. Limits are lowering. That noose tightens until the economics make sense for companies walking into trillion-dollar IPOs.
I started calling the hidden part ignorance debt. It compounds... and later is always the most expensive time to deal with it.
The Math That Doesn’t Add Up
When I paid $200 flat every month, no single request cost me anything I could feel.
I’d open Claude Code, fire off a prompt, and never think about what it ran on. Fable 5, the frontier model? Sure, let’s go. More tokens. Don’t think about it...
There are limits buried in those plans. But how many of us can see where the line is? (I know some of you run multiple accounts and watch the line closely. I get it.)
Per-token billing is the honest way to sell AI. You pay per request, not a flat monthly plan. You see the cost. You decide if the request was worth it. You can say, “that one was worth it, that one wasn’t.”
But someone has to foot the energy bill. AI is not free. Things cost money to power. GPUs cost money to power. That’s just the nature of the universe.
So if you’re not paying enough to cover the bill and a little profit for the provider, who is? With the flat plans? Probably not you.
That flat cost erased and buried everything for me. I stopped thinking about the cost of a request. I stopped thinking about which model I was using. I opened the app and kept working.
The Skill That Atrophies
The money was the smallest piece. The bigger debt was a skill I didn’t know I was losing.
The skill of judging whether something I paid for was worth it. In a free market, that anchoring matters. Otherwise you float free from reality, because everything just “doesn’t cost any money.”
The day Fable 5 originally got pulled, I built a full web directory in about six hours. Web scraping, API keys, SEO and AEO, the works. A year earlier that project would have eaten a week. I gave it no skills, no subagents, no context. I just said, “Hey, do it.”
Why could I do that without giving it any instructions? Because I wasn’t paying for the inference. There was no moment for me to check the value of what it was creating.
That thing could have cost me $500. $1,000. Has it made me that much since then? No. When Fable 5 built my whole directory, the output belonged to Fable 5, not me.
That directory cost something to run. I couldn’t see the price. And we need to be able to perceive that value and weigh it against the cost. Using tools that subsidize the inference weakens that skill, even if it’s subtle.
The Lock-In You Can’t Feel
That leads to the third part of the debt: switching costs.
You start with something because it’s convenient. You stick around because it’s cheap. Then you build around that tool. Not the tool itself, the habits that go with it. And eventually the convenience becomes the reason you don’t want to leave.
That’s vendor lock-in (stuck because leaving is expensive). The neural pathways get stronger until that’s just what you do.
Compounding is quiet and patient. Every week you stay, leaving gets a little more expensive.
And later... was exactly where I kept putting it on my to-do list.
A Debt Is Not a Debt Sentence
A debt is payable. You pay it down on purpose, a little at a time, and every payment makes you stronger instead of weaker. The bill coming due was never the part that scared me. The part that scared me was how long I went without looking at the statement.
(Believe me. I’ve gone deep in credit card debt. No shame. I take pride in having clawed out.)
The debt was never about Claude. Claude was simply my lender. They’re running a good business. It’s a free world; they’re free to sell a product and I’m free to stop buying it.
The real debt is the convenience you couldn’t replace if you had to: the file you couldn’t rewrite, the skill you used to have, the reality you used to know.
Run it through the only question that matters in this game: do I own this? Not the invoice. The capability. If the answer is no, you’re very likely the product, even if you’re paying.
Three Payments
So how do I pay it down? Three payments.
Payment one already happened: the exit. I canceled Claude Code and went exclusively to OpenCode, an open-source harness. There’s also Agent Zero (which I haven’t used as much lately, but it’s a viable option). The point of these harnesses is that you can switch to any provider, any model, and the only thing that changes is the output. I now pay for every token I use, and every session I can see what it cost me.
If you want the full breakdown of what the open stack does that Claude Code doesn’t, and where it honestly falls short, I wrote that comparison up separately after running a real job on the open stack for $9.
Payment two is the audit. I’ve been going through my workspaces file by file, workflow by workflow, asking two questions of everything: Can I explain this? Can I move it, use another model with it? If the answer to either is no, it goes on the debt ledger and gets paid down before anything new gets built on top of it.
Payment three is the report. Every couple of weeks I make insights I feel like I should have had a year ago. Cheaper techniques. Things I can finally explain. And I’m only learning them because I stopped using Claude Code.
The open-weight models are getting faster, smarter, and cheaper while I do this. DeepSeek v4.1 Flash is dirt cheap, good, and fast. I never would have said that about any open-weight model a year ago. No chance.
That’s why now is probably the right moment to start mapping your exit. Because those models are good enough that the trade is real, not theoretical.
What It Costs You
I don’t miss Claude Code at all. That’s not true. I do miss it. (I do.) The way it compacts long conversations and hands off context between agents. I’ve vibe-coded some simple replacements in OpenCode, and I’m not getting the results I like yet...
There are things about Claude Code I miss. But I don’t need it anymore. I miss the subsidized inference for building random fun things. But I don’t need to build those, and they distract me from the actual business.
What the debt was actually costing me was agency. The ability to leave. The ability to choose. The version of me that can do the work without asking permission from a $200/month plan.
That’s worth more than any tool. And it carries into the rest of your life.
I’m still super ignorant. But it feels good to have started making that decision.
Your Ledger Is Already Open
If this is vibing with you, somewhere in your setup right now there’s probably a convenience you don’t fully understand. Maybe a file you couldn’t rewrite. Maybe a tool you can’t imagine living without.
Consider that your ledger. Dive into it. See if you can figure it out. Pay that debt back.
This is about the knowledge, and paying the debt down so you’re not caught off guard when the terms change.
If you want the full cost picture before you compare anything against it, I broke down what Claude Code actually costs separately. And if the real problem is a drawer full of subscriptions with no system behind them, I wrote how I cut my own AI bill from about $200 a month to roughly $20 by matching the model to the difficulty of the task.
The architecture that makes any of this possible is the harness: the instructions, skills, and files you build around a model. The part you own. The model underneath is a rental. Rentals get swapped.
That’s the belief we carry at the Academy, and it’s the instinct this whole essay is built on. If you want to learn to build AI workspaces you actually own, workspaces no provider can take away from you and that you can change on a whim, that’s what we do at AI Captains Academy. Twice a week we get on calls. We run workshops. We share what is working. It starts at $83/month or $497/year with a 7-day free trial.
We use AI. We love AI. But we do not want to be owned by it.
Questions people actually ask
- What is ignorance debt?
- Ignorance debt is the hidden, compounding cost of using subsidized flat-rate AI. The flat subscription hides what each request actually costs, weakens your sense of whether a request was worth it, and locks you into the tool through habit. Like financial debt, it is quiet and patient, and later is always the most expensive time to deal with it.
- Is a flat-rate AI subscription bad?
- Not inherently. The convenience is real, and while you are learning or building hard things the flat plan removes a meter-watching anxiety that genuinely helps. The problem starts when the flat cost erases your sense of what the work costs, and when you build habits and files you couldn't move to another tool. Use the subsidy; don't let it own your stack.
- How do I pay down ignorance debt?
- Three payments. First, the exit: move to an open-source harness like OpenCode so you can switch providers and models freely. Second, the audit: go file by file and ask "can I explain this, can I move it?" Anything that fails goes on a ledger you pay down before building new things on top. Third, the report: the cheaper techniques and explainability that only show up after you leave.
- What's wrong with subsidized inference?
- Nothing is wrong with it, and taking advantage of it while you can is rational. The catch is that subsidized things change terms: limits lower, prices rise, features gate behind higher tiers as companies move toward profitability. If your whole workflow depends on the subsidy, the day it changes is the day you discover the real cost was never the invoice.
- Do I need to cancel Claude Code?
- No. This isn't a call to cancel anything. It's a call to own the capability underneath whichever tool you use: the instructions, skills, and files that survive a model or provider swap. If you can't explain or move your own setup, you carry the debt whether you keep the subscription or not. I left because I wanted to feel that ownership; your path can look different.
- What's an open-weight model?
- An open-weight model is a model you can download and run yourself, rather than one locked behind a single company's API. DeepSeek v4.1 Flash and GLM 5.2 are current examples: cheap, fast, and good enough that the gap to frontier models has narrowed sharply in the last year. You reach them through an open harness like OpenCode or Agent Zero.
