My AI coding costs kept going up until I was spending about $650 a month on GitHub Copilot. That’s a lot of money for tools I mostly use to write and review code, so I wanted to see how far I could get without paying anything.
For a month, I switched to Qwen 4B running locally on my MacBook, an M4 Pro with 48 GB of RAM. It was free. After that month I went back to Claude, and today I use both. This post is about how I got there and how I split the work now.
How the bill gets that high
I don’t think my situation is unusual. When an AI tool is part of almost everything you do during the day, you stop thinking about individual requests. Each one feels small. The monthly total is the first time you really see what it adds up to, and by then it’s a habit.
Seeing that number was enough to make me question whether I needed a paid model for every task, or only for some of them.
A month on a local model
The MacBook made the experiment easy to try. Apple Silicon uses unified memory, so the CPU and GPU share the same pool of RAM. With 48 GB, there’s plenty of room to run a small model on the machine itself. A 4B model is small by current standards, which is a big part of why it’s practical on a laptop.
The obvious benefit was cost. Running a model locally has no per-request charge, so there’s nothing to keep an eye on. I think that changes how you use a tool. You don’t hesitate before asking a quick question or trying something a second time.
In my view, small local models are good at simple, well-defined tasks. Autocomplete, quick questions about syntax or an API, short explanations and small edits are all things a model that size can handle well enough. For a lot of everyday coding, that covers more than you might expect.
Where small models fall short, in general, is multi-step agent work. Planning a change across several files, deciding which tools to call, and keeping track of context over a long task is where larger hosted models are clearly better. That’s not a criticism of small models. They’re built for a different job.
Why I went back to Claude
After the month, I went back to Claude. Part of the reason was practical. I now work on a Windows mini PC, and the local setup I’d been using was on my MacBook.
The other reason is that Claude is good at complex work, including things beyond writing code, like generating images and logos and making Figma designs. A small local model isn’t built for that kind of work.
How I split it now
Today I use Claude for complex work and a local Qwen model for everything else. For me, that’s the best of both worlds. The expensive model only gets the work where it makes a real difference.
What I like about this split is that it isn’t a compromise in either direction. I’m not forcing a small model to do work it isn’t suited for, and I’m not paying frontier prices to answer a syntax question.
What I’d suggest
If your AI tool bill has crept up, start by looking at what you’re actually using it for. A lot of it is probably small, routine work that doesn’t need the most capable model available.
Then try a small local model for a while on those tasks. If you have a Mac with Apple Silicon and a decent amount of memory, it’s easy to try. You’ll learn where it’s good enough, and where it isn’t, much faster than any comparison chart could tell you.
Keep a hosted model for the hard work. Multi-step agent tasks are where the bigger models earn their cost, and trying to save money there tends to cost you time instead.
The main thing I took away is that one model for everything isn’t the only option. Matching the model to the task kept the tools I rely on and stopped me paying top price for work that didn’t need it.
Join the discussion