Notes.

While I'm perfectly happy with the work Opus 5 has been doing for me, I'm giving Codex and Chat GPT a try as well. So far, it feels pretty clean, without all of nutso verbosity, and the work feels good as well.

Opus 5 launched today, so I've been using it. It seems to be doing pretty well so far. Of course, that's what I thought about Opus 4.8 for a few days, and then 15M tokens in one afternoon put me off the model for a while.

We'll see...

I do not like the word "footgun". Sonnet 5 just gave me the bad news about it creating a situation where code that was pulled into a foundation package to be shared amongst applications required sharing multiple copies of the file across each repo. A real "f**tg*n".

I'm enjoying working on code with Sonnet 5. It does high -quality work, is relatively fast, and cheap as far as usage goes. I guess my only gripe is sometimes it makes up verbs (fullizes???).

Having now worked with Claude Code for a few months building several apps using Laravel, I think they are a natural fit. The code is usually pretty clean, and it utilizes the feature set very well. Having spent some time on Reddit and heard stories that isn't always the case for code written by AI.

I'm a little less impressed with the JavaScript and Blade templates, which usually end up with too much hard coding, scripts in the blade templates. But still, I'm pretty happy with the quality of the code overall.

After that first weird blip it's doing pretty good, and it's cheaper than using Opus 4.7 (We don't talk about 4.8 around here).

Sonnet 5 dropped today, so I decided to let it work on the latest Beta fix wave for Carrel. Already on the first ticket it tried to create a Python script to fix one line of a test PHP file, so I don't know, man.

Anthropic enabled Fable for my Max 5x account for a couple weeks, so I gave it a try. Used it to do an optimization check for a slow loading admin view on one of projects, and also to do a comprehensive review of whether or not the code matched the architected model. Am thinking of a couple more things before it gets taken away in a little over a week.

All in all, I like it! But I feel like I don't know exactly what to give it to try to untangle.

Opus 4.8 shitting the bed update! It ate usage like Cookie Monster eats cookies. On top of the weirdness, I can't trust it. Today I've gone back to Opus 4.7, which seems to be working out fine.

Looking on Reddit, it's clear I'm not the only one who has had weirdness with 4.8. Also, Anthropic reset usage earlier today, which maybe is because of this whole deal.

Just all part of the fun, I suppose!

Finally ran into a situation where Claude Code running Opus 4.8 kinda shit the bed. The code still seems to have been written and modified fine, but getting there was awfully weird. At one point it saved a memory note that I was unavailable due to my dad visiting (which was not true and was never something I said anything close to being true to Claude) and a whole bunch of other weird stuff. Claude Code has mostly been rock solid for me since I started using it back in April, and I had no issues like this with Opus 4.8 since I started using it a few days ago. All part of the fun of using Claude, I suppose.

I remember trying to spec out a code creator that worked in written English back in the early 2000s. I wasn't deep enough into Computer Science to be thinking of AI, which of course wasn't anything close to what it is now. But working with Claude Code feels like the realization of this dream.

Now that I've been developing several projects on my own, I've gotten really acquainted with Laravel's seeders. They really help create test data really fast. Couple with Claude Code, they are incredibly powerful. The painful days of feeding test data by hand are over!