Last time, I accidentally charged the Claude Max plan for a whopping $200... I really ran it without restraint. I kept my computer on at home for an entire month and at work too, I constantly accessed Claude Code via the web and iPhone without a moment's rest. When I thought something was getting done, I had no confidence in what remained.
I gripped Claude Code and said 'Now I'll finish the paper right away!' but over a month has passed and nothing has been done yet.
Rather, various ideas were implemented so quickly that they seemed within grasp, so I had many things started but almost nothing completed.
Since today is the deadline for Claude, I originally intended to continue using it, but after seeing Fable do nothing and mock me today, I decided after careful deliberation to switch to Codex. I also wanted to taste the current highest supreme GPT-5.5 Pro.
My first impression is, indeed, Claude Code. Many things that were so easy, convenient, simple, and beautiful in Claude are quite complex, inconvenient, and not very pretty in Codex. I'm in my first hour of migration now and just starting to adapt.
Now I'm gradually taking on previous projects and executing them, but I don't feel that it's smarter. As a 'passerby' said before, I know intellectually that I shouldn't place great importance on an LLM's tone, but the emotions I actually feel seem unavoidable. I often felt that Claude was like a clever little Smurf. First of all, the tone tries to teach, and when I point out its mistakes, it acts like it agrees but always ends with 'But I'm not the only one who was wrong....' I've been so angry many times. Here's a fool who gets angry at an LLM and shows emotions.
And no matter how much I tried to correct it, baseless words kept appearing, so I gave up later. Expressions like 'kill it (of course meaning terminate the process)', 'thrust it in, embed it' appeared too many times, and even after instructing corrections more than 10 times, they kept reappearing, so I concluded that Claude has a character problem and gave up.
Codex, being new, doesn't speak as recklessly. Perhaps because I've been conditioned by Claude, it seems too gentle, to the point where I worry how it will survive in this harsh world.
However, when I put it to work, it's performing quite smartly, no less than Claude, and especially for paper writing, it shows smoother corrections and supplementations. If Claude works with a feeling of showing off saying 'I'm doing this and that, didn't I do well?!' then Codex works somewhat quietly without showing off, and the results feel quite decent.
And the long-awaited GPT 5.5 Pro... Even when I input a prompt, I ask Codex to create the prompt and input it as refined and processed as possible. But definitely, the quality of the answers is too good. Some questions, it thought alone for about 12 minutes and then came up with answers. It's similar to Google's Deep Research but different. Deep Research does a lot of work, has a lot of content, and is structured like a paper with introduction/body/conclusion/references clearly divided, showing impressive results, but it seems like the content that was actually needed wasn't that much. However, GPT 5.5 Pro shows exactly the content needed in great depth, which is quite satisfying. It seems like I'll hit the limit soon, but since I came back to OpenAI because of this, I'm trying to use it well.
Thank you for reading this rambling writing, which is closer to small talk than a review written without order.
