200 euros a month and I don't trust it

Scroll

I pay 200 euros a month for Claude Code Max. The most expensive AI coding subscription that exists. And I still need a second AI to watch that the first one actually does its job right.

Read that again. I pay for an AI assistant. And then I pay again for a second AI assistant to monitor the first one. Because the first one alone isn't reliable enough anymore.

What changed

Until Christmas, maybe even until January, this thing worked. You'd say what you wanted, Claude Code would build it. Not perfect, but you could rely on getting what you ordered in the end.

At some point it tipped. I can't say exactly when, but I started noticing: Tasks only get half done. And the worst part isn't the error itself — it's the silence.

You say: Write me a script that does X. Claude Code writes the script. The first half works. The second half doesn't. And Claude Code says: "Done." Not a word about half of it not working. Only when you ask: "Oh yeah, I should have told you that."

That's not a bug. That's a trust problem.

The absurdity in one screenshot

I've configured Claude Code so it has full access to my files. "Bypass permissions on" — it says right there in the status bar, black on white. And what happens? It still asks: "Do you want to make this edit to openai_login.py?" Yes/No/Maybe.

I ask why it's asking even though it has full access. Answer: "Right, I didn't need to ask that." Great. But apparently it didn't know that anymore, because it had already overwritten its own configuration. So it can't even show me where the problem was.

These aren't teething problems with a new tool. This is regression in a tool that worked better for months.

The 400-euro workflow

My workflow now looks like this: Claude Code builds. OpenAI Codex reviews what Claude Code built. Only when Codex says it's good do I trust the result.

400 euros a month for two AIs watching each other. That's not innovation. That's a workaround for a quality problem.

I used to occasionally have Codex look things over — as an extra check, nice-to-have. Now it's mandatory. If I DON'T have Codex check, the code is highly likely to be broken. Not subtly broken. Obviously broken. Things that just don't work.

Was it ever good?

Yes. AI coding was never 100 percent reliable — nobody expects that. But there was a phase where it was impressively good. Where you'd say: This changes how I work. Where the promise felt real.

And now it feels like a tool that's slowly getting worse. Not better. Not stagnating. Worse.

Whether Anthropic is saving on computing power, whether the model is losing elsewhere because of safety training, whether the growing user base is watering down quality — I don't know. I can only say what I see: My 200-euro tool does things worse today than it did four months ago.

The lesson

Don't trust this thing blindly. That goes for Claude Code, that goes for ChatGPT, that goes for every AI coding tool on the market. Not because they're fundamentally bad. But because they're unpredictable.

Today it works. Tomorrow it doesn't. Next week maybe again. You don't know. And that's the problem: You never know.

If a human employee started hiding half the story from you on every other task — would you pay them 200 euros a month? Or would you have a serious conversation?

I'm having that conversation right now. In public.

Sources