Why I don't spend more than 30$ on AI coding tools

I honestly don't understand how one can spend more than 30$ per month for AI coding tools. And stay sane.

I have been using AI for engineering tasks over the last few years, from a more sophisticated autocomplete to fully autonomous agents completing my sluggishly written instructions. I use them for many hours a week.

Yet, I hardly spend more than 20-30$ per month on these tools. Why?

I think one should make a difference between vibe coding and agentic engineering.

For vibe coding I currently use MiniMax M2.7. It costs me 10$ per month and basically runs on near-infinite tokens. It performs worse than the current frontier models, but I don't really care. I currently use it in Claude Code and with "bypass permissions mode" on. If it fails, I just prompt again. So far, it has found a solution to almost anything I wanted.

What do I vibe code? Scripts! Scripts to automate everything. Analytics, data analysis, content creator tools, blog post and newsletter translators, literally anything that doesn't need specific engineering principles and is essentially throw-away code. These scripts sometimes use the OpenRouter API for access to specialized models which sometimes cost me another 10-20$ per month.

But again: 10$ a month for a coding agent that never reaches a limit. Has probably saved me hundreds of hours of manual work already.

And what do I engineer with the help of agents? Sofware! I am currently working on a quite complex architectural refactor of nightride.com. It will allow for a lot of new features and automate the platform even further. Not my favourite work, I'd rather ship features.

And boy, is it mentally straining. Much more difficult to stay in the flow.

Claude Code (the real one, now) is my buddy here. It helps me streamline and double check my architectural decisions. We work on a big architecture document together which tracks a list of issues and PRs – and their mutual dependencies. I let Sonnet 4.6 with medium to high thinking effort write the issues, based on my ideas and thoughts. I let it challenge them. I let the same model with low thinking effort execute them, create PRs, and merge them into the develop branch, where I manually test it. I always merge into production myself. I invoke a couple of self-defined skills that run tests and style checks, and sometimes I let specific agents impersonate a coding guru and critisize the work. I regularly skim the code – out of curiousity, but also out of necessity. Over time, I developed a feeling for what works out of the box and what doesn't. And it keeps me sharp (I hope).

Yet I am still on the Pro subscription – for 20$ a month. I usually implement 1-2 PRs in the five hour usage window. More than enough.

Two reasons:

1. More agentic engineering would burn me out. It's tempting, it's addictive, the next feature only a prompt away. "Usage limit reached" is my reminder to go out for a walk, to think, to reconsider. Because this is what produces the moat of nightride.com. Not a machine. And shipping features without thinking is only creating technical debt without any added benefit.
2. Sonnet 4.6 is just good enough for these kind of tasks. I have used Opus 4.7 once: by accident. I don't think it would make the architecture much "better" or produce a bit less code smell. Maybe it would oversee one thing less, okay. I frankly don't understand the hype.

My theory: These engineering agents will have quite soon reached a point where "good enough" is really good. The price competition will only increase, while AI companies struggle to meet demand and cover costs. It might all end very soon with a big crash.

I see the current time as the golden time of agentic engineering: When it was still affordable and fun. The time to profit from it is now. But don't let it burn you out. More than 3-4 hours of agentic engineering per day just does not make sense. Engineers need to find a better way to spend the rest of their day.

And don't tell me that somebody who "burns" through thousands of dollars in tokens produces anything useful that will scale and be maintainable.

The only thing that is burning here is mental health.

Do you want to know more about my setup? Contact me. Happy to share more in a call.