Claude got so dumb, so I stopped relying on it alone

Soooo.... I became a content creator, turned founder again, and lately someone who spends his evenings arguing with Xcode about why a button won't align. I did not expect the last few months to go the way they did.

I went from building a flagship MVP, to a custom dashboard built entirely for my own workflow, to now fumbling around in Xcode and finding out the hard way that building native Mac apps in Swift is no joke.

Claude AI and Claude Code did about 90 percent of the heavy lifting on all of it. Without them, I wouldn't have been able to build any of this on a simple $20 Pro subscription, let alone migrate every one of my sites off WordPress, including this one.

I'm usually the guy who sticks to one platform. Not just because switching is a hassle, but because I try to live by digital minimalism: fewer tools, used with intention, instead of hopping between apps for the sake of it.

Claude forced me to find alternatives

I don't know what the fuck has been going on with Claude lately, but it got noticeably worse. When Fable 5 launched, it was rough.

Sonnet 5 on high effort used to be my default, and nine times out of ten it could one-shot my prompts in Claude Code or nail changes to a mockup on the first try.

Somewhere along the way, that stopped being true. Some sessions got so bad I had to close my laptop and remind myself I was talking to a model, not a person who should know better.

A few of the worst moments from the past two weeks:

• After 119 attempts and iterations, including dozens of reference screenshots and exact instructions on what to change, Claude still ignored most of it.

• It would randomly bring up things out of context, pulling in details from other projects that had nothing to do with what I was working on.

• It started changing things I never asked it to touch.

• I asked eighteen times to match a footer's font color to the top navigation. Hex code included. Screenshots included. Code attached. It ignored the request every single time.

It felt like managing a dumbass intern who has seven diplomas framed on the wall but can't turn on his own computer. It's that bad.

For a moment, I genuinely considered canceling my subscription, leaving Anthropic with a big fuck you and moving on.

People noticed too, not just me

Turns out I wasn't imagining it. When Fable 5, Anthropic's flagship model, launched in June 2026, the backlash was immediate: the launch drew heavy complaints about high token burn, an undisclosed tendency to quietly weaken its own output on certain AI research tasks, and a mandatory 30-day data retention policy.

After Anthropic paused and then restored access to the model, Reddit users reported the restored version still felt weaker, describing it as routed through stricter safety systems and falling back to Opus 4.8 far more often than expected.

And it's not only a Fable problem. Part of the broader "Claude got worse" complaints trace back to developers calling generic model aliases instead of pinned versions, meaning an app can silently switch to a different model whenever Anthropic updates that alias.

Either way, I wasn't the only one who felt like Claude had taken a step back.

I made a shift, but Claude stayed my primary driver

Then I thought about the Spacemail outage. It taught me something: a single point of failure, or just not diversifying at all, can cause a lot of unnecessary stress.

I'm usually careful about hedging and thinking through platform risk, but Spacemail was a case where I only had myself to blame for not being diligent enough.

So I split things up. Claude seems to lose the fucking plot when I load it with too much scattered context, so I made a short list of what I actually need help with:

• Coding and development

• Polishing blog posts, grammar and sentence-level edits

• Imagery for Pinterest first, other platforms later

• Inspiration for new posts or follow-ups to something I already wrote

• Research and talking points I can use as a starting reference for writing

• Stories written for English B1 to B2 learners

And here's where I thought:

I shouldn't put my entire workflow, memory, or every vertical I'm working in, into one model. It's not that I'm worried about getting locked out of my account, if that's even a real risk. It's that spreading specific context and workloads across different tools just makes more sense.

I assumed this would cost me two more subscriptions. It didn't cost me anything extra.

Where Gemini and ChatGPT help me more than Claude

Gemini's free-tier Flash model is, in my experience, far better at picking up on what I mean without needing detailed instructions.

For a while now I've fed it everything about the blogging side: SEO, sitemaps, the kind of content structure that fits each of my blogs.

If I tried the same thing with Claude, it wouldn't get past the first two requests. It struggles to properly read a URL or a published post and give me an honest, neutral read on it.

That's where Gemini has been genuinely useful. It gives better guidance on content strategy, understands my tone of voice, and gets that this blog is more of an outlet than something structured for SEO.

Ten out of ten for Gemini here, two out of ten for Claude.

I still like cross-referencing with a second AI, which for me is ChatGPT. These days the only thing I really ask it for is a list of points to argue a position. Things like:

• Make me a list of points that argue whether UGC creators are ignoring blogging

• Make a list of arguments for and against the idea that blogging is dead

I never copy any of it directly. If you scroll through my posts, it's obvious I don't give fuck all about SEO or structured content. But ChatGPT's output usually makes sense.

It helps me frame points I already understand but didn't know how to put into words. Less like being taught something new, more like being reminded that I already knew it.

Where this leaves me

After testing all three side by side, Gemini and ChatGPT outperform Claude for the content and research side of what I do.

Claude is still my default for daily development and shipping, including the Mac app I'm still mildly terrified of finishing.

Once things scale further, I think it'll be worth paying for a dedicated subscription per use case instead of throwing everything at one model and hoping it holds context.

Claude getting noticeably worse in areas I couldn't quite believe forced me to rethink my setup, and honestly, that turned out to be a good thing. I can keep coding and shipping centralized in Claude, lean on Gemini more for the content side, and use ChatGPT as a third, neutral check.

That said: Anthropic, whatever happened with Fable, fix it. Claude got noticeably dumber in ways that were hard to ignore, and it's the reason I stopped trusting it with everything.