Moving away from things is difficult. Changing your workflows and working style is even more so.
That was probably why I had put up with the Anthropic coding plan for such a long time. Didn’t help that I had a pseudo-ethical backbone in me that rooted for the underdog in the whole AI race. Well, the underdog at that time, at least.
It didn’t take long to reach the breaking point. My days were filled with dread about reading the final outputs of Opus 4.8.
The ginormous text blobs with circular wording that, whichever way I read them, made no sense whatsoever. The ‘caveats’. The ‘honest truths’. The ‘load-bearing seams’. The ‘that’s on me’s. My eyes became like SIMD-accelerated regexes scanning for meaning in paragraphs actively trying to hide it.
The Breaking Point
And it didn’t happen overnight either. The Opus models were actually known for great writing style, with 4.6 being the absolute peak Opus model both in intelligence and writing style. Things started becoming weird with Opus 4.7 and went downhill rapidly with 4.8.
It’s like boiling a frog or water under the mat. Not immediately noticeable, so you can’t just put your finger on it, but things were getting worse and worse.
We Tell You What You Can Do
The final breaking point was receiving a mail from Anthropic about the restrictions on using it in non-interactive mode.
Yes, my $200 sub can apparently not be used with my own harness which I have engineered over the years to be not token-hungry, perfectly tailored for my experience and workflows.
But the die was cast at least a few days earlier. This was later admitted by Anthropic, but I had to find out the hard way.
You cannot do any kind of ML work with Anthropic models.
They do not refuse it outright, of course, but they do something even worse. They silently drop the effort or reasoning quality for those prompts marked as ML. The model would make absolutely boneheaded basic mistakes and apologise profusely when pointed out, while repeating the same mistakes over and over again.
I was doubting my sanity around ML work when they finally announced the policy.
Codex, Cold Turkey
I cancelled my Max subscription the same day and moved to Codex cold turkey.
I thought the transition would be disruptive, especially because I had such bad experiences with GPT models a while back.
Porting my custom harness to use Codex was attempted with Opus, but the sub ran out midway, so I had to use GPT-5.5 to complete it.
And I’m glad it ran out midway. Because the contrast was night and day.
No More Expletive-Driven Development
No longer was I constantly feeling the need to punch my laptop reading the blobs that Opus 4.8 slopped out.
GPT-5.5 was way more thorough and methodical. It’s terse to a fault and… just… did things?
It’s difficult to describe it, but one noticeable outcome was that I was no longer doing ‘expletive-driven development’. I could walk away and come back with absolute confidence that it would follow the prompt to a fault and wouldn’t take a shortcut or go over my work.
These thingies got personalities
Working with Opus was like managing a constantly lying, lazy employee with a drug habit and schizophrenia combined into one. You never know if the thing is gonna do its work or stab your codebase in the heart.
Working with GPT-5.5 was like working with a neckbeard with a PhD. Absolutely no taste in frontend or design. Wears the same belt and suspenders every single day, but completes the work exactly as told and agreed on. And hunts down corner cases and bugs like it’s nothing.
It does struggle with working with defective workflows or broken dev setups because it takes docs at their face value. Opus, on the other hand, never had any such issues. Well, probably being so used to taking shortcuts and lying to the users, it probably expects the same from all the codebases it works on!
One thing that is immediately noticeable about Codex is that OpenAI has a lot of compute. The responses were ridiculously fast and it wouldn’t slow down even with a high number of projects running in parallel. No more constant retries or 529 errors streaming on my screen.
I was happy. The only regret I had was that I didn’t switch earlier. The work was getting done faster; I actually made releases rather than stay depressed reading the outputs.
The Great Fable FOMO
Then came the Fable announcement. The world was singing its praises. It’s like there isn’t a single tech blog or article or forum not awash in its praises.
Usually I would FOMO back into a new model release, but surprisingly I couldn’t care less. I’m finally shipping.
It’s like I finally found the most reliable high-mileage car that takes me to my destination and I no longer feel shorthanded on the development side.
Well, Fable didn’t last long. Bans swiftly came, then the nerfing, then the limits. I was glad I stayed out of that constant drama that is Anthropic.
Then Came Sol
I was excited for the new model release because of how bad GPT-5.5 was at designing anything remotely looking like a UI.
Sol was better but not as good as even previous versions of Opus in terms of design. Still, it was a welcome improvement.
The first two days, Sol crunched on my backlog and things were slow. It went in circles trying to hunt down every edge and corner case in the product. It wrote a zillion tests that were overengineered.
My flashbacks of working with Opus models came back. The work went in the garbage bin.
It’s kind of expected with any new model, but Sol seems to have been aimed solely at reliability and one-shot performance rather than iterative development.
And this is why every engineering organization should and must have their own harness (or a fork of an existing one) tuned to their workflows. The difference between a tuned harness working with Sol and an untuned one is day and night.
What the Subscription Buys
The limits across Codex are quite generous, but, as most Codex users know and appreciate, the random resets they provide—sometimes to celebrate their milestones and other times as a peace offering for outages and bugs—are quite helpful too.
Well, you can’t depend on them for regular work, but you certainly can use them for completing those long-pending one-off tasks like codebase reviews, POCs, etc.
A Codex subscription comes with a separate limit for ChatGPT chat, which might not sound like much in the grand scheme of things, but it lets you research on different designs, requirements, product ideas and even personal non-work-related stuff without worrying if you will be cut off midweek due to limits.
While all these are great, the single best thing I continue to appreciate about them is how good their models currently are. If, hypothetically, they never release their next model and I’m stuck with Sol, I’ll be happy to just use it for the next few years.
If you are not on Codex, do yourself a favour and get on it.