You’re a developer. You want to build with AI. You want to inspect every system prompt, see exactly what the model is doing, and walk away anytime you want. No vendor lock-in. No black boxes. Just you, your code, and a model that obeys you.
Then you look at the pricing. Claude Pro: $20/month. ChatGPT Plus: $20/month. GitHub Copilot: $10/month. All cheap, all seamless, all heavily subsidized by the model providers. And the BYOK open-source clients—the ones that let you bring your own key and inspect everything—they feel like a hassle. The API costs seem higher, the setup more complex, the performance slightly worse.
So you take the easy path. You click “Subscribe.” You think you’re saving money.
You’re not saving money. You’re paying with your future freedom.
I’ve seen this play out again and again. A team starts with a subsidized consumer plan because it’s cheap. Six months later, they’re deeply integrated into that ecosystem. Their workflows depend on proprietary features. Their codebase has been trained on opaque models. Switching? That would cost more than the subscription ever saved them.
This is not an accident. Model providers are deliberately subsidizing consumer plans to create dependency and lock-in. They’re making the ‘true cost’ of independent BYOK setups seem artificially high—so high that you don’t even try. Transparency becomes a luxury good, reserved for those who can afford to pay full price or are willing to fight the friction.
The most expensive AI is the one that owns your pipeline.
Look at the HN thread that started this. A developer asked: “How to use agents via API cheaply?” They wanted to inspect everything, avoid vendor lock-in. They mentioned Deepseek and Kimi as efficient and cheap. The top comment? “Outside Claude and Codex, I use Kilo Code and it’s good so far.” Note the phrase: “outside Claude and Codex.” Even the alternatives are framed as exceptions to the subsidized norm.
Here’s the twist: the independent path is not actually more expensive. It’s just more transparent. The subsidized plans hide the cost in your future lock-in. When you BYOK, you pay per token—but you own the whole pipeline. You can switch models, audit prompts, and build truly portable agents. The upfront cost is higher, but the total cost of ownership is lower—if you value autonomy.
But most developers don’t compute that. They see the $20/month price tag and think it’s a bargain. They don’t see the hidden cost: the gradual surrender of control, the creeping dependency, the slow erosion of their ability to leave.
Independence is not a premium feature. It’s the default that subsidized plans are designed to take away from you.
So before you sign up for that cheap API, ask yourself: who really owns your code? If you can’t switch providers in a weekend, you’re not building—you’re renting. And the landlord is about to raise the rent.
FAQ
Q: Aren't subsidized plans like Claude Pro actually cheaper for individual developers?
A: They appear cheaper on the surface, but they lock you into a proprietary ecosystem. The cost of switching later—rewriting integrations, retraining workflows, losing data—almost always exceeds the savings. BYOK setups cost more per token but give you full portability and control.
Q: What's the practical implication for a team starting with AI agents today?
A: Choose your tooling with exit in mind. Use open-source clients that support multiple providers, inspect system prompts, and allow you to swap models without rewriting code. The few extra dollars per month now will save you thousands in migration costs later.
Q: Isn't this just paranoia? Many developers use subsidized plans without issues.
A: It's not paranoia—it's history. Every major platform play (App Store, AWS, Facebook) started with low prices to create dependency, then raised them. AI providers are following the same playbook. The developers who don't plan for exit are the ones who end up trapped.