I tried Claude Code because I was frustrated with Codex

2026-10-02

For a long time, I was genuinely a big fan of the Codex models.

And I wasn’t quiet about it either.

I praised them a lot because, for a while, they deserved it. I was using them for real work — building, debugging, auditing, working through architecture — and they were performing really well.

Then September happened.

I don’t know exactly what changed, and I’m not going to pretend I do, but the experience got noticeably worse for me.

The models started feeling slower. They also started feeling… dumber.

And Luna especially started driving me fucking insane.

Luna could be incredibly good one moment and then completely lose the plot on something that felt straightforward. I’d give it context, explain what I wanted, and somehow end up spending more time correcting the agent than actually benefiting from it.

The latency made it worse.

When a model is great, you can forgive waiting a little. When you’re waiting and then the answer isn’t even good, the whole experience becomes frustrating very quickly.

It reached the point where I started wondering whether I should just try something else.

So I thought: fuck it.

What if I buy a Claude Code subscription and see whether it’s actually better?

I did.

And oh my God.

Opus

The first thing I noticed with Opus wasn’t some benchmark number.

It was that I could trust it.

I could give it a messy problem, let it explore the codebase, and it would often understand what I was actually trying to accomplish instead of just mechanically completing the most obvious interpretation of my prompt.

That difference sounds small.

When you’re building real software, it isn’t.

I started giving Opus serious work on Musika and Sekiwa: architecture decisions, audits, nasty bugs, things where understanding the surrounding system matters just as much as writing the code.

And I kept having this feeling:

Oh. This is what I was missing.

I wasn’t fighting the model as much.

I could think about the problem instead of constantly thinking about how to make the agent think about the problem.

That was huge for me.

But there was one obvious problem.

Opus is expensive.

Really fucking good — but expensive.

So I still couldn’t treat it like an unlimited worker. I had to think about where using Opus was actually worth it.

And then I tried the new Sonnet.

Then Sonnet surprised me

I expected the normal compromise.

Cheaper model. Faster model. Good for normal coding. Bring Opus back when things get serious.

Except Sonnet was much closer to Opus than I expected.

I started giving it work that I previously would have reserved for Opus.

And it kept handling it.

Then I tried lower effort.

Still good.

Not “good for a cheaper model.”

Just good.

There were tasks where I genuinely had to remind myself that I wasn’t using Opus.

That’s when my entire way of thinking about these models started changing.

I don’t care which model wins a benchmark

At this point, the question I’m interested in isn’t:

Which model is the smartest?

It’s:

At what point can I actually feel the difference?

Because that’s what matters when I’m building.

If Sonnet can handle the implementation, debugging and reasoning I’m doing every day, I don’t need to burn Opus usage just because Opus is technically the stronger model.

Opus can sit there for the problems where I actually need it.

Deep architecture.

Difficult audits.

Really nasty bugs.

Problems where Sonnet starts going in circles or I can tell that something is missing.

Everything else?

Sonnet is becoming very hard to argue against.

The funny part

I didn’t move to Claude Code because somebody convinced me Claude was better.

I wasn’t chasing benchmarks.

I moved because I was frustrated.

Codex had been genuinely good to me. That’s why the September experience bothered me so much in the first place.

I knew how good these tools could feel when they were working well.

I just wasn’t getting that experience anymore.

So I tried Claude Code.

Now I’m using Sonnet for work I would previously have thrown at the strongest model available, and keeping Opus around for when I actually need the extra intelligence.

That’s probably the biggest compliment I can give Sonnet.

It didn’t make me think Opus was bad.

It made me realize how rarely I actually need Opus.

And considering I came here because Luna was pissing me off?

That’s a pretty funny way to end up becoming a Claude Code user.

Yes, no shame — I use ChatGPT to clean up my grammar. The thoughts are still mine 😂

← All notes