Why I'm Not Paying for ChatGPT

September 2026

Everyone serious about AI assumes the same starting point: you pay a subscription. ChatGPT Plus, Claude Pro, Gemini — pick your flavor, sign up, move on.

I don't pay for any of them. I run everything locally, and it's not the reason you think.

This isn't a privacy screed. I'm not wearing a tinfoil hat about my data floating through a third party's API. The reason is simpler and more stubborn: I don't want my thinking tools held hostage by someone else's pricing page.

I'm not against OpenAI. I'm against building a dependency on a service I don't control, and cannot budget for. Especially when an emerging market has shown a volatility for consumption-based pricing that makes any planning near impossible.

Here's the moment it hit me. I had been happily building against a hosted model when the provider changed its pricing — not by an amount that broke the budget, but by an amount that made me stop and recompute: how much does this actually cost me to operate for a month? How much will that same 30 days cost me in a year? Most importantly, can I even budget for how to handle it if it changes this week? Privacy is important; being able to keep a customer-facing service in operation is not far behind. When the frontier operators go to war, prices change overnight. Sometimes that's a win... sometimes not. For me, the answer to every one of those questions lives entirely in someone else's hands, so I spent an afternoon standing up a local alternative and I never looked back.

Now my whole stack runs on hardware I already own — three machines on my network, a handful of local models, one small Windows box with a GPU doing the heavy inference. The total monthly cost is roughly the price of the subscription I'd otherwise be paying, except there's no meter.

Running local models means I can experiment without watching the meter. Failed prompt? Delete it. Wrong model? Swap it. Need to try eleven different approaches to a problem because one of them might be better? Go. No incremental cost, no commitment, no per-token anxiety. The economics stop shaping the questions I'm willing to ask. This is a technology in flux; tightly coupling yourself to a model is on par with betting big on Betamax. And if you get that reference, you're old.

I'll be honest about the tradeoffs. The latency is worse. Hallucinations are more frequent. A local model won't match the very top of what the hosted frontier offers, so you iterate through options, trading some time and effort for cost. I'd call the quality on most of what I do about 85% of a flagship model — and some days it's below that.

But here's the thing I keep coming back to: 85% that I control beats 100% that can change its pricing model tomorrow. Control shows up in ordinary ways. My models don't get deprecated under me. My prompt history isn't a feature someone can quietly remove. When the tool is mine, every improvement I make to it is an asset that stays on my shelf, not rent I pay to someone else's.

I'm not saying subscriptions are wrong. For most people, they're the right call. Cheaper, simpler, best-in-class. That's a legitimate trade.

But if you're building a career on these tools — if your judgment and your working method are getting tangled up in them — it's worth asking who actually owns the thing you think. For me, the answer is: I do.