The $40 Subscription Is Dead: Why I Switched My Whole AI Stack to Open-Source Models in August

From February to July, I paid $40 a month for tools that rationed me. $20 for Claude Code, $20 for Codex — both closed, both cutting me off at 5 hours, while my business was bleeding from a tourism crash and needed every hour I could get.

In August I switched my entire stack to open source. Today my daily drivers are GLM 5.3 Flash, DeepSeek V4 Flash, and Qwen 3.8 Flash, running on the DeepSeek Harness — my own customizable agent runtime — with OpenCode and Command Code as my front-ends, about $20 a month in tooling. The models everyone called unusable are now my primary drivers, and honestly, they kick the tires off the premium model I used to pay for.

What the Closed-Source Months Actually Looked Like

Before February, I had no AI subscription at all. My daily driver was an old WordPress website and Facebook content, and every inquiry got answered by hand — manual tasks, every day, while my ranking sat so low my own listing was not in the top 10.

Then two things happened at once: tourism in Baguio crashed because of the Iran war, and agentic AI went mainstream. My business was not doing well, so I learned AI to fix my own business — Claude Code for building, Codex alongside it. $20 and $20. That is what the whole recovery was built on: one man, $40 of subscriptions, and a business that needed to be saved. The chatbot I built in that period runs to this day.

But the walls kept appearing. Five hours of use, then a hard stop. Want more capacity? That is another tier, another price. And the prices kept climbing. I was renting my tools from companies whose business model is that I never own anything — not the model, not the limits, not the price.

The August Switch

The switch was not ideological. It was two frustrations stacking up: the 5-hour limits always landed at the worst time, and Claude Code and Codex were getting more expensive while doing it. So in August, I tried the thing everyone said was a downgrade.

I was shocked. The three Flash models — GLM 5.3, DeepSeek V4, Qwen 3.8 — were not just usable on my daily driver, they were genuinely capable. Content creation is really good right now. What open source could not do before, it does now, and the gap closes every month while closed tools raise prices to close their own gaps with shareholders.

My Stack Today, Job by Job

One model per job, chosen on cost and capability — no romance about any of them, same as it ever was. Here is exactly what runs what.

The old stack is not fully gone — Gemini still makes my reels at $5 a month, because it is the best tool for that one job. That is the point. I am not anti-closed-source, I am anti-bad-math. The day a closed model is the best value for a job, it goes back in the stack. The day an open model beats it, the swap takes minutes, because nothing in my setup is locked to a vendor.

The Math: $40+ and Rationed, or ~$20 and Unlimited

Here is the honest comparison, not the fantasy one. Before: $40 a month in subscriptions, plus the chatbot API on top, plus 5-hour walls that cut my worst month into pieces. Now: about $20 a month in tooling — $10 OpenCode, $10 Command Code — the same $30-60 chatbot API at my volume, and no walls anywhere.

So the fair claim is not "everything became free." It is this: the subscription layer was cut in half, the usage limits disappeared entirely, and the models got more capable at the same time. For less than $100 a month, a business owner has a complete daily driver — you have your own employee. And because the cost is so low while the output compounds, the return on my own business multiplied to the point where a listing that was not even in the top 10 now shows up consistently and stays fully booked — packed even in the rain. That is my own measured result, not a promise, and Google Analytics backs it: up 2,103 percent.

What Open Source Unlocked That Renting Never Did

The cheaper part is obvious. The bigger part is what nobody warns you about. When you own the harness, you can add anything you want to it — my marketing skills, my blog pipeline, my chatbot QA loop, all of it lives in my setup, not behind someone else's feature gate. When a better open model ships, I swap it in the same afternoon and nothing breaks, because my tools are not married to a vendor's roadmap.

And the models keep coming. GLM, DeepSeek, Qwen — three competing open-source families, improving against each other in public, every month. Closed tools improve too, but their improvements arrive as invoices. Open improvements arrive as Tuesday.

The Mistake to Avoid When You Switch

Budget blindness almost got me twice. The first time was on premium models — run a high-end model on every message and the bill explodes at volume; done unknowing, it can cost more than an employee. The second trap is subtler: free or cheap is not a strategy either. Check what the best model is in terms of quality and quantity for the job, consider the cost at your real volume — if inquiries jump fifty thousand a month, the wrong pick blows the budget at exactly the moment you are winning — and decide deliberately. The win is not "open source is free." The win is "I know what every token costs, and I chose it."

Who This Story Is For

Small business owners still sticking to traditional methods because AI seems too expensive, too technical, or too risky: I was you. A transient owner whose own listing was not in the top 10, running everything manually on WordPress and Facebook. The tools were closed and rationed, and I built anyway — because the business was bleeding and waiting was the more expensive option.

The lesson is not "run what I run." The lesson is that the cost of building with AI fell off a cliff in August, and it fell for everyone. The owners who start now compound — the site, the content, the chatbot FAQ, all of it stacks month over month. The owners who wait are not standing still; they are falling behind businesses one-tenth their size. I am happy with my business, and I can help you with yours — that is literally the job now.

The One Thing to Remember

In February I paid $40 a month for tools that rationed me. Today, open-source models run my whole business — inbox, content, landing pages — for less, with no limits, and they beat the premium model I used to pay for. AI got cheaper and more powerful at the same time, and the gap is widening every month. Chinese models win right now. We do not know what the future holds. So do not marry a brand — run the math every month, own your harness, and build while the building is cheap.