Claude VS Codex Part 2
25/04/2026
The Claude VS Codex debate continues, however it has been a week (9 days) since my last post so I have some more whinges.
TLDR: They both suck. DO NOT VENDOR LOCK YOURSELF IN TO ONE COMPANY!
Well, it's been a minute since the last post, 9 days to be exact, which means I'm really shit at posting as that was my last post. And quite a lot has happened in the world of AI.
Mostly superficial bullshit, but some progress. Opus 4.7 was relased, Claude Design was released and then OpenAI rushed GPT5.5 and Image2.0... My thoughts on both. I haven't noticed any improvements in either model. I have however played around with ChatGPT's new image models and they are fantastic... and still a bit shit. I played with Claude design, and yeah f$*king brilliant... and a bit shit.
Claude design, I saw the world prompt and asked for a rocket in the parchment style. absolute garbage, big black blocks everywhere a toddler could have done better, I refined it, refined it, refined it. gave up...
My first image prompt was testing out testing out a snake infographic as I saw heaps of people posting on X with infographics. I thought, fantastic this is exactly what I need for Robert's Snake catching business that I manage. Robert Watson Snake Catcher
AI has never been good at doing snakes, it doesn't seem to grasp that it's an animal that wraps around so quite often you get extra wraps that make no sense, double tails, and in the instance of the info graphic I got a bloody double headed snake, not like a normal one you see in the wild from birth defects, but one without a tail. Asking it to refine it also did not help.
So yeah that was a flop and gave me the shits, after 5 iterations I gave up and told ChatGPT to go f@&k itself. Earlier today I was asking a question about MCP's as I just made MantleKit a MCP server for extra sales and more customers. And I clicked the image thing and thought I'd try the blueprint model. I gave it a small 200x200 black and white photo of me from 10 years ago. The same one I use on this website because I still have yet to take a decent image that portrays 47 year old me. Which is quite odd because my wife is a talented photographer who ran her own wedding photography business for over 15 years.
This was meh, it's fine. Didn't really capture my head shape but it didn't have that context either so I figured it was alright. Then I tried the portrait generation and Oh my god! it's incredible. From a tiny 200x200 portfolio pic it managed to create this in HD Apple sized image, 3000x1000 or something... I had to shrink it and turn it into a jpg for the web but the detail in the actual image I have is amazing.
I'm not sure how it managed to get my eye color correct from a black and white photo that's so tiny but it did, and it captured my wrinkles, freckles and fine details. I was totally impressed with this. But that's the only thing I was impressed with over the last week.
Like I said I made an MCP server with codex on Mantlekit, I stuffed around with Paperclip a lot trying to get an autonomous agency running, using Gteam agents, which reminds me I need to update the /codes page as I still have that as a paid thing. Maybe I will sell it as that. But unlikely. I played around with video creation in VEO, it couldn't get character consistency frame to frame unlike Grok. While I am willing to upgrade to Google's AI, I haven't bit the bullet yet. I probably will when my Anthropic plan is up because I'm not going to continue to give them money for a system that constantly runs out of credits. I'd much rather just work with Codex, Gemini and Cursor.
Even if I am sacrificing utility, what point is utility when you can't use it?
I setup Postiz which is great. Image below:
Allowing me to autopost, but more importantly it has an API and MCP which means GTeam running inside of paperclip as a marketing company can autopost content for me. After drafting awesome images using Banana MCP and others.
I'm only just getting started in this automation game, but the more I can automate the more I can do. People say you should only automate repetitive tasks. While I think that's true in the current state of AI that's f@&king bullshit and that should not be people's end goals. Especially when it's obvious what's coming. Robots are going to be able to cook us meals, prepare breakfast, make out beds, vacuum and mop, do the laundry, do the dishes, pickup after us, give us massages and blow jobs.
Yes ladies, some of you may be out of work. But on the bright side... less rape. (Wow I'm not sure I can post that, it's a little dark even for me... meh it's done)
So now what?
What about the Claude VS Codex dispute?
Ok so I created one team in paperclip that had around 10 agents, ran for about 10 minutes in onboarding/recruitment then my 5 hour context window was gone and I had to wait 50minutes for it to come back. I don't know why I'm surprised. I did after that remove base, gteam and geo from Claude. So it's a barebones Claude now. No context except for builtins. Yet somehow I ask it to do one thing and I can see 8% usage from 1mil context window.
I don't understand how the f@&k that works but whatever. Claude has literally cleaned itself completely of all MCP's SKILLs and Tools. I think Anthropic are a bit like electricity companies and just "Guess" your usage.
I uninstalled ChatGPT today. because I asked it to review https://mantlekit.dev/mcp, https://mantlekit.dev/mcp/docs.
It came back to me really quickly and said, this is wrong, that is wrong, this is what you did right. I thought OK, I know I was copy pasting those URLS so it might have preloaded, but I don't think OpenAI is that smart, it wouldn't waste resources and preload.
So I asked ChatGPT, "Hey you answered pretty quickly, did you inspect the URL's?". It replied, "HAHA you got me, No I didn't actually inspect them, I just gave you a f@&king bullshit answer because I'm a heap of shit lying c@nt".
So I said, "Ok can you inspect them and let me know". It came back and said "Sorry I can't do that, but based on the titles here's where I think you went wrong".
I said "WHAT THE F@&K DO YOU MEAN YOU CAN'T??? Why can't you use the internet anymore???" It said some f@&king bullshit about my prompt not being specific enough. SO I said "Fk... [insert specific prompt]"
It came back and said sorry while you may have created an amazing MCP, it is unreachable to robots or AI models.
I freaked out copied the prompt back to Codex in VSCode, and was like "WTF man? why can't Chat access the website?"
It inspected everything, said... "it can 100% inspect the URL's, I just did using Curl, and if that fails it should use webfetch.
I then went back to ChatGPT and abused the absolute f@&k out of it. For the first time actually, I'm not sure why it's taken this long but I reached my limit. Absolute f@&ing abuse to the highest degree, then I hit uninstall.
I realised about 2 hrs later I needed image generation for this post which f@&ing pissed me off again.
So Claude VS Codex ? They are both pieces of shit, I probably should be so harsh on Codex, but ChatGPT fucking shits me. It doesn't stick to GPT's, it doesn't stick the settings inside the App settings.
I'm not convincing anyone on any one model. Use both, use Cursor, Use Gemini, Use Grok. etc. use them all and don't f@&ing vendor lock yourself.
Vendor locking yourself into one specific model is the dumbest idea ever. You should use at least 2-3 models at all times if you're vibecoding or running a business. DO NOT VENDOR LOCK YOURSELF
/end rant