Overview

I’ve been exploring Claude for over a month now and have been using it more regularly to the point that I ran out of my weekly session limits. This got me into a situation where I was tossing up between paying an extra $80 for the month to get the $100 max plan, or using a different AI provider for the remainder of the week. But after reading around and seeing that I might not have my weekly limit updated even after upgrading my plan, I decided to try a Codex.

Initial impressions

As I’ve been doing more coding, I opted to try and use the Codex CLI right away. On an initial impression, I was a bit underwhelmed by it. Compared to Claude Code, it feels like the Codex CLI has a bit more to go. This was on version 0.142.2.

I installed the Superpowers skill, which I had to do manually unlike Claude. In Claude, there was a plugins install command that I used to install the Superpowers skill. With Codex, I had to run a Powershell script to copy the files and install it. Granted, it wasn’t that difficult but I was hoping there would have been a plugins install command.

Using skills and gates

With the Superpowers skills, I also noticed that the Codex CLI ignored all the gates that Superpowers asked for. Usually when I ask for something, a plan is drafted and I would normally have to review and approve it. With Codex CLI, it just bypassed all of that and assumed that I approved. When questioned about it, it recognized that it bypassed all of the gated checks which was confusing.

This kind of defeated the purpose of using the skill with Codex since the interaction and feedback I provide is one of the more useful aspects of the Superpowers skills.

GPT

In terms of the LLM itself, I didn’t really notice much difference between it and Claude. It was able to help me design and build an application after I learned to plan with it more. I do think that Claude and Opus was able to provide me with better results when I asked for UI designs. I’m not sure if that was because my prompts were wrong or the way the LLMs were trained. When I asked for 8 different design palettes for my application, I liked Claude’s designs a bit more.

I think Claude was able to use a frontend-design skill that I had installed which prompted the LLM to not use the standard fonts. I had to explicitly state that with GPT before it used the skill.

Pull Requests

When it comes to pull requests, Claude definitely wins on description. I’m still a bit unsure why Codex’s descriptions for the pull requests are so simple. Claude’s ones are usually very descriptive and explains the changes in more details. Codex’s pull requests tend to be bullet points that are short and straight to the point. At the moment I like Claude’s descriptions more.

Model comparisons

When it comes to token usage, I think Opus 4.8 uses more tokens. I’m tempted to use Sonnet more but I do like using the most expensive model because I don’t really have a lot of time to fluff around. GPT5.5 on extra high works pretty well and at the moment, it’s really hard for me to do comparisons because they do the job that I ask of it. I’m just unsure if the AI harness is the one that needs a bit of work.

Conclusion

It’s getting more interesting to explore AI harnesses and see the differences. I think I’ll try out Kimi next as that’s been getting really good reviews recently.