Cancelling Claude, it's the right time for me now because stupidity runs off the charts.

It's Saturday. I wasn't planning to write anything extra today. Then I sat down for what should have been a fifteen-minute pass on a file. A simple adjustment, and six rounds later I had nothing to show for it except a burned limit and the specific kind of anger you get when you've explained the same thing four different ways to something that keeps nodding along.

It got so bad that I wanted to complain in this chat session and see how it responded.

I've been paying for Claude since May 2026. The moment I started my vibe coding journey. You can clearly see how my GitHub went from zero to 100.

Anyway, this month I'm cancelling it, and I hope someone from Anthropic will read Reddit forums, this blog post, and a ton of other similar cases.

This isn't one bad session

I want to be clear about that, because "AI had a bad day" is a boring post, and that's not what this is. In fact, I still praise Claude. It opened a whole new world for me, as a non-developer.

It's every single request now. Not most. Not the complicated ones. Every fucking one. I ask for a change, I highlight the exact lines, I circle the thing, I state explicitly what needs to happen and what must not be touched, and Claude does something else. Then I point it out. Then it does something else again.

I could write three pages of specification, every detail spelled out, edge cases included, and it still wouldn't land it. I've tested that theory. Repeatedly.

The length of my prompt has stopped correlating with the quality of the output, which is a nice way of saying the input barely matters anymore.

Read also: Claude got so dumb, so I stopped relying on it alone

Earlier this week I was working on my BETA invite. Something that should be simple turned into the most complicated shit ever.

Another day it took 119 attempts to iterate. I counted, because at some point counting was the only thing keeping me from throwing the laptop.

One hundred and nineteen iterations and it still got it wrong. Then a few days later I came back to fix the fallout from that, and it fucked that up too, having understood nothing about what it built the first time.

That's not a bad session. That's a pattern.

I tested every model, every setting

Before anyone tells me I'm holding it wrong: I tried all of them. Opus 5 on high, on default, on low. The other models, same thing, same settings, same spread. I moved between them, looking for the one that still worked.

There isn't one. The failure is consistent across the whole lineup, which honestly makes it worse. If one model were broken, I'd switch models. Instead, I'm switching providers.

"You're right, and this is on me"

This is the part that gets me, and tilts the fuck out of me for real.

When Claude fails, it apologises. Beautifully. "You're right, and this is on me." Perfect contrition, immediately, every time. And then the next response makes the same mistake.

An apology in a chat window costs nothing. It doesn't give me back the round. It doesn't give me back the limit it just burned to produce the thing I have to reject.

It's a nice sentence sitting on top of a failure, and the sentence keeps getting more polished while the failure stays exactly the same size. I'd rather it said nothing and got it right.

Read also: Claude AI Is both my greatest tool and my biggest frustration

$965 billion

Anthropic closed a $65 billion Series H in May at a $965 billion post-money valuation. That's above OpenAI. They filed confidentially for an IPO a few days later. The reporting puts their revenue run rate around $47 billion.

They are a motherfucking, functionally, trillion-dollar company, and from what I read, they're aiming higher.

You would think a company at that scale could ship one or two models that reason properly and don't fall apart on a straightforward request. That's the whole fucking product. Not the valuation, not the funding round, not the IPO paperwork. The product is: I ask for a thing, and it does the thing.

I'm not owed a perfect model because they raised money. But I am paying for this, and the gap between what that number implies and what shows up in my terminal has gotten hard to ignore.

So I'm cancelling, and trying Codex instead

My original plan was softer than this. Keep Claude Code running with Cursor on the Pro plan, move content creation and ideas over to Gemini, use ChatGPT as a neutral third party for verification and double checks.

Diversify without paying for three Pro subscriptions. Stay on free tiers elsewhere until I understood the limits well enough to know what was worth paying for.

That plan is fucking dead, because the part I was going to keep is the part that stopped working.

So: cancelling Claude. Running Codex and ChatGPT for a month.

I've heard good things about Codex, and I want to see for myself. What pushed me over is that the free tier of ChatGPT has one-shotted requests that Claude missed six to twenty times. Free. Against a paid subscription. That comparison is not subtle.

On the content side, it's already settled. Gemini outperforms Claude by a wide margin for creation, audits, feedback, and structural thinking. Also on the free plan. I stopped pretending otherwise a while ago.

Luckily for me, this comes at the right time, because today my subscription was about to renew, but that's not going to happen.

September is the right month to do this

The timing isn't random, and it isn't just spite.

September is a content month for me. Writing and recording are the priority. On the product side, it's BETA invites for my flagship MVP and QA checks, which means no real development work. The stakes for tooling are as low as they're going to get all year.

The Mac app I'm building is still experimental, so it's a fine place to test drive something new. The note-taking app for language learners is even safer. I have the repo, nobody's using it, and if a new model makes a mess, I can throw the mess away.

If I'm going to run an experiment on my entire development workflow, this is the month where a failed experiment costs me the least.

And I get a blog post out of it either way, because I'll be writing down what I find and comparing them properly.

I'll email Anthropic anyway

There's no real point. Their support is notoriously fucking bad, and I don't expect a reply that isn't a template, let alone compensation.

I'm sending it regardless. Partly so it's on record, partly because complaining into a blog post and never telling the company feels like the wrong order of operations. If they respond, I'll say so here. If they don't, that's its own kind of answer.

One last thing

I gave this a long runway. Longer than I should have. I kept extending the benefit of the doubt because the tool was genuinely good once, and I kept assuming the next update would fix whatever had broken.

It didn't. It got worse. Claude got so fucking stupid, and I mean that plainly, not as a joke. The thing I used to hand a half-formed idea to and get something usable back now can't follow an instruction I've written down and pointed at.

So the subscription goes. Maybe I come back in October, and it's fixed. Maybe Codex turns out to be worse, and I eat this post. I'm fine with either outcome, because right now the only guaranteed result is the one I've been getting all week, and that one is nothing.