AI as Collaborator: Why I'm Skipping the Bill and Going Local

There's this endless back-and-forth in developer circles that I can't shake, mostly because I sit in an awkward spot relative to both sides of it, not fully buying either extreme.

On one side: the people who say AI will replace programmers, that you don't need to learn to code anymore, that you can just describe what you want and ship it.
On the other side: the ones who say real developers don't use AI, that it's cheating, that the code it produces is garbage and the people relying on it don't understand what they're building.

Neither of those positions matches what I see when I look at the developers I actually respect. John O'Nolan, for example, is using AI while building Alcove. Or seasoned professionals at big companies, using it to accelerate work and speed up tasks they already fully understand and master. The pattern isn't AI instead of skill or skill instead of AI. It's deep understanding of the problem, with AI as a capable collaborator for the implementation.

That's what I'm aiming for. I committed to those JavaScript, Node.js, and React courses because I want to understand what I'm building. Not as a formality, not to check a box, but because shipping code I can't explain or maintain isn't something I'm willing to do. Especially when I'm eventually building tools that other people might use.

And I'm also not going to pretend to have some ideological objection to using AI while I learn. If something I need already exists and AI helps me understand it or build a version of it faster, I'll use it. The distinction I actually care about is this: Am I using AI to avoid understanding something, or to extend what I already understand? The first produces things I can't maintain, the second produces things faster. This distinction is what really matters.

The bill (almost) nobody talks about

Here's what the "just vibe code it" crowd tends to skip over (or they just flex about): the cost.

Cursor starts at $20/month for the Pro plan. But try to use heavier and most recent models or drag out sessions, and those credits vanish quickly. Continue to push it and you're at $200/month for the Ultra tier.
OpenAI Codex follows the exact same numbers, except it's not a separate product, it's bundled into your ChatGPT subscription: $20/month on Plus, $200/month on Pro. Claude does something similar: the $20/month Pro plan bundles both Claude Code, the coding agent, and Cowork, which was launched in January 2026 and applies the same agentic approach to general knowledge work (file management, research, documents, etc). One subscription, several tools. Up to $200/month if you need the higher Max tiers. The pattern is the same across all of them: the base plan sounds reasonable, but for more intensive work, the next available plan costs 10x more.

There's also a "pay-as-you-go" (Extra Usage) option across all of these that I'd strongly advise anyone learning to leave disabled, unless you enjoy surprises on your credit card bill. But if you really need it, you can set a monthly limit so you won't exceed more than you need.

Google Antigravity launched in November 2025 and is currently free for individual/personal use during its public preview, with generous rate limits on Gemini 3.1 Pro. That's worth knowing. But paid subscribers already get higher limits, which tells you exactly where this is going. And it's Google: impressive on paper, sure, but not even an option when privacy matters more than convenience to me or anyone with the same values.

The pattern across all of these is the same: the editor is often free or cheap, the model is the product. Stack two or three of these together, which plenty of developers do, and you're looking at $200 to $600 a month before you've shipped anything.
For a developer with a clear return on that spend, the math makes sense. For someone in my position - still learning the fundamentals, no product yet, with a long-term goal of building free and open-source tools - it doesn't add up at all. You'd be paying hundreds of dollars/euros a month to build things you're giving away for free.

For the past year or so, I've been using Claude and Grok because they're tied into my agency workflow, and the subscription's cost more than pays for itself with the time it saves me on client work. But I'm not going deeper into that ecosystem than I already am. The direction I want to go is the other way.

Where I'm actually heading

My long-term plan is full local AI. On my hardware, with open models, no outside servers, no fees or caps. And crucially, no data leaving the room.

I've already installed Ollama and pulled down Qwen3-Coder-Next, which was released in February and is specifically designed for coding agents and local development (or so they say). For the interface, I'm eyeing Jan, this open-source ChatGPT stand-in with over 5 million downloads and 40k GitHub stars, seems like a good option. Instead of terminal chats with Ollama, Jan layers a clean UI over your local models, all running on your machine. It's built in public, matching the open ethos I'm chasing elsewhere. That said, I've only scratched the surface. That's the honest answer and the only one I can give right now.

I should also clarify what I mean by "local". There are tools like OpenClaw, the trending open-source personal AI assistant created by Austrian developer Peter Steinberger (recently hired by OpenAI), which blew up from a side project to nearly 250k GitHub stars in a matter of weeks. Rather than giving you a new interface to learn, it works through the messaging apps you already use, like WhatsApp and Telegram. You install it on your own machine (It even spiked Mac Mini shortages in some places), bring your own API key, and it does support local inferance through Ollama. But the whole approach, an always-on assistant with deep access to your system, executing tasks autonomously while you're away from the desk, is a different thing and a bigger commitment than what I'm talking about. I just want a model on my machine I can think alongide, and that can assist me when I really need it. The autonomous agent stuff is not part of my plans.

What I do know is that open models are genuinely getting good and fast. A year ago the gap between local and commercial models was significant enough to matter for most tasks, but that gap is narrowing. From where I am right now, (learning code, talking through approaches, catching errors) a well-configured model is likely more than capable to fullfil my needs.
And the per-query cost is zero.

And honestly, it's not just about being consistent with everything else I'm building, although that matters a lot too. We're at a point where the big companies already know more about people than they probably should, and most of us keep voluntarily feeding them more (work stuff, personal stuff, things that probably should stay private and we wouldn't have the courage to say in public). I'm not here to tell anyone how to handle that, it's not my place.

But my work, my learning, the way I think through problems... That's mine. I'd rather keep it that way, and that isn't up for negotiation.

When I finally get the time to set Ollama up with Qwen3-Coder-Next and give it a real test run, I'll write about it here. If the workflow and results hold up and the model does what it promises, great. If it turns out to be more trouble than it's worth, I'll write about it too, and we'll go from there to find a valid alternative.

May The Code Be With You! 🚀