Back to Windows: Why CachyOS Didn't Stick as My Daily Driver

Hey guys, I know I haven't posted in a while. Here's what I've been up to.

For the last several months I used CachyOS, and Linux in general, as my daily driver. I've now gone back to Windows 11.

Linux Has Come a Long Way

I want to be clear about this first: Linux has made giant steps in compatibility with Windows programs and games. Proton is very good now. Most of my library just worked, and I barely had to touch any settings. If you've written Linux off for gaming because of how it was a few years ago, that's outdated.

It's Not Enough, Though

Good compatibility doesn't make a system stable, and stability is what made me leave.

I kept running into memory problems. Usage would hit 100%, and instead of the system handling it, everything crashed. The only fix was rebooting, every single time. It happened often enough that I couldn't rely on the machine for anything I cared about.

What I'm On Now

Windows 11, but debloated. I did a clean install of 25H2 and applied FSOS-X with AME Wizard, which strips out the stuff I never wanted. So far it's been stable.

I'm not done with Linux. I'll probably try it again later, but for now I need a machine that stays up.

AI Models and Tools

I've also tried basically every AI model under the sun over these past months, so here's where I landed.

Opus 5.5 from Anthropic is the one I'd suggest everyone try. It's a big jump over every other model I've used, and right now it's the frontier. It's also cheap per task, which I wouldn't have said about Opus before.

Usage is better too. At $20 a month you actually get real usage out of these models. With Opus 3 and 4 you'd get a few calls per 5-hour window and then you were done. That's not the case anymore.

I also spent time with Codex and its models, Luna, Astra and Sol. Luna is the standout for me. In terms of cost per task, it's very efficient.

Sonnet 5.5 deserves credit too. It's quick, straight to the point, and smart even at low thinking levels, which I didn't expect.

Moving Away From Pi

In my earlier posts I talked a lot about Pi. I've moved away from it, and I'm using Claude Code and Codex now.

The reasons are simple:

• Better integration with the models they're built for
• More control over the actions the agent takes
• Non-frontier models without a dedicated harness aren't as relevant as they used to be, so I don't see the point of Pi for now

Pi is still a great piece of software, though. If your needs are different, it's worth a look.