Claude Code Regression - Can we stop writing AI CLI with AI CLI itself?
Can we talk about absolute tragedy of Claude Code?
Where should we start? Probably from the beginning. Small intro into this: https://www.youtube.com/shorts/eRov-sFcf1s — or you can just believe that Claude Code was initially created by accident, for devs inside Anthropic. Throughout 2025 it became more and more popular and at the end of the year it became the ultimate building tool. I am a long running Cursor user on the 60$ plan and I still think it’s a perfect balance, but I really wanted to stay in sync with all other developer world. I mean, did you see how many people shipped stuff just using Claude? And I was stuck in my mind prison with ideas that I can’t build because of time.
I am a fan of Cursor, hater of Codex CLI and Theo subscriber.
Cursor was not my first AI tool for coding — it started with Supermaven, continued with early Windsurf (I think it was in beta). I tried Cursor multiple times and landed on it after a lot of tries because I wasn’t satisfied with the other products. By the time GPT-5 dropped I was also handed the GPT Team plan sub which gave me access to the first version of Codex CLI, and even though I used it a lot — I just hated it. Everything was bad at that time for me: handling the context, conversations, multichats, bugs, speed sometimes. I switched to Claude Code in December ‘25, which is pretty late, and it is really groundbreaking. I paid 200$ to use Ralphy, Opencode and build what I want, test things and just explore. Truly groundbreaking moment. The video that kept me happy spending 200 bucks on AI was none other than: https://www.youtube.com/watch?v=-5LfRL82Jck. But I was also not on the side of “Anthropic is the saint company” — I knew that they had issues, and Theo mentioned this a lot, which was a glimpse of the future for all of us.
First nail in the coffin — We are The Ones and Only.
Not a secret by this time that Claude Max sub is very generous, for their own reasons of course — easy to get users and become a dominating company in the AI world, and maybe people even start referring to AI not as ChatGPT but as “Oh, ask Claude.” But one morning changed everything: Claude Max sub is no longer usable externally or in Opencode. I am a user of Opencode as well — I love the UI, how they handle chat history and a lot of things. Every software has its issues, and Opencode is no different, but every hour you’ll get a pop-up about updating it so just update and enjoy. Also it’s OSS, so you can just add and fix stuff yourself — why not?
Thousands of bans, drama and weird answers from Anthropic itself. Let’s talk about Sonnet 4.6: https://www.youtube.com/watch?v=M-pkXr-qqII.
Seriously, this video shows a lot of what I am talking about here so I don’t want to repeat the same information. The only thing I can say is that locking users to one tool is not a great idea when you are not sure that you will be better forever. OpenAI was not in a great spot by that time so they had the advantage of Opus 4.5.
Second nail in the coffin — Bun
Can I just say it? I love Bun with my heart but Anthropic has a very weird specific view of the world and company business. I believe it’s great to support a tool to build better things — I almost believe acquiring Bun would help them build better tools — but I also believe they should stop digging the grave for themselves. The main issue with Claude Code is not having Bun, it’s Anthropic being the owners and builders of Claude Code.
Third nail in the coffin — YOU BANNED!
Roughly after two months of Claude Coding I went from “Wow, AI really can do so much — Opus is so great” to being banned, probably because of using Ralphy or because I messaged Opencode with the max sub one time to check if the ban was really valid. The worst policy I’ve ever seen is not explaining anything about the ban — no explanation, just ban and fill up a Google form and we will answer you, which I did fill and never got any answers. I continued using Claude Code with Max sub because they refunded me and I just used another account, which is insane. You know this money glitch I never heard anyone talking about? Buy a sub, try to use something Anthropic hates and get fully refunded even if you’re on week 4. This is so bad it’s not even funny, and the reason for this support policy is because they’re banning users not for doing some very dark stuff or breaking the models — just because you wanted a TUI and not the CC CLI.
Fourth nail in the coffin — OpenAI
Yes, while Anthropic thought they were the only ones in the AI world, big Sam cooked and shipped a Codex App for macOS. It was faster, much more UX-friendly, easy handling of multiple chats for different projects, showed the thinking process — and I just love using something with a UI, to be honest. Felt so much easier. They also shipped pretty good handling of MCPs, Skills and etc. But the time has come to talk about…
Opus 4.6 vs Codex 5.3
No, I won’t compare these models because I , again , completely agree with Theo’s point — use both. While I’m writing this, Codex app testing the mobile app and Claude Code testing another project, so what is it then? I just hate using Claude Code, this is it. It became much more buggy and Opus became slower. For minutes I was not able to tell if the model had actually bugged out and stopped making progress, or if it’s just how Anthropic handles Thinking.
Do not mind my message, I accidentally sent it. Unfortunately, this is not the only bug I had, and I will again refer to Theo.
What now?
Of course, I could advise Anthropic to stop writing CC with AI and a lot of other obvious things — but who am I to do that? Only time will show the future. What I can suggest now is: go get a GPT sub while it’s 2x limits, grab Claude Max sub if you have the money, and ship. I’m sure I’ll switch to the Codex 200$ plan next month, or 100$ if they ship it. I’d still use Claude but for 20 or 100, ideally 50. As of writing this on February 22nd, Codex offers a Spark model that is exactly the same but just super fast, and in the case of Opus they’ll give you 50$ for the worst fast model. I just hate it — why can’t I use your models with a 200$ plan and have to pay extra for 1 million context or a fast model? Genuinely bad.
My fellow developers — use Codex App on Mac, wait for Codex on other systems, try the Codex CLI (it is much better now!) and enjoy using CC while we can abuse at least the usage token of their plans.
Happy shipping.