articles about Anthropic's Claude code, and how I use it for programming, and various techniques I use to make it more efficient. However, for the past two weeks I've been experimenting a lot with OpenAI's Codex and I'm seeing much improved results compared to Codex a few months ago.
In my opinion, Codex is equally good for most tasks and has the advantage that, in most cases, it is faster than Claude Code and that it is better at doing what it is asked to do and not doing other tasks (which is a problem I have experienced with Claude Code).
In this article, I'll be discussing my experiences using OpenAI's Codex for advanced coding tasks and other areas of the program, as well as some techniques I use to improve Codex performance.
Why use OpenAI Codex?
First, I want to cover why you should use OpenAI Codex. It is worth mentioning that the price of Codex 20x Max subscription is the same as Claude Code. The only difference is the quality of the results produced by the model and how effectively it can complete the tasks.
Considering that I program every day, it's important for me to stay up to date with the latest coding models and to always try new and upcoming models, like GPT-5.5, to see if it works better than my current setup.
I just started using Codex with GPT-5.5 about two weeks ago and just used it for some of the real world projects I was working on. This is important, as I believe that using coding models in testing activities does not really test the strength of the model, nor is it a valid and complete test.
When I used it on some of the more complex projects I was working on, I was very impressed with the results. In my opinion, Codex was very efficient in completing other tasks and completed very quickly. In addition, I had the impression that Codex was better than Claude Code in doing the exact tasks I asked it to do and not changing other things in the code. Actually, the problem I've had with Claude Code a few times is that I ask it to complete a task, and it usually completes that task, but it also changes other things that I didn't want to change.
It is fair to say that this is very balanced. On the other hand, you have Claude Code's approach, which is to give the model more freedom to make decisions about what to change, which can lead to the model changing parts of the code you don't intend to change. On the other hand, you have the Codex method, which only changes what the user asks you to update. This, on the other hand, can have the disadvantage of leading to states in the entire codex because it is not updated, because the Codex does what it is asked to do and nothing else.
Some techniques I use to prepare the Codex
In this section, I'll cover some techniques I use to make Codex work better out of the box. I will cover my setup and some techniques.
My setup
First, let's close my setup. I'm using fast mode in Codex at the moment because I don't hit my limits very often. However, if you hit your limits, you should consider turning off FastMode or getting another Codex account.
In addition, I use more high logic when I use the editing mode and high logic or logic when I use the normal mode, and I use GPT-5.5, of course.
I also gave Codex access to the Playwright MCP, which is how I can access my browser and perform actions there. This works very well, for example, with OpenClaw bots, which I'll cover in the next section, as well as in-browser logging and testing features introduced by Codex. As I have mentioned in many previous articles, allowing your coding agents to test their work greatly improves the performance of these coding models.
You can read more about this in my article below:
How to Make Claude Code Validate Your Own Work
Lastly, I also use YOLO mode with Codex, where I give it, or allow it to do, any action within the folder it's working on. In my experience, borderline code models, like Claude Code and Codex, are not prone to making bad mistakes like deleting production databases or similar, and will usually warn you before taking irreversible actions.
Furthermore, I also believe that if you set up your codebase and infrastructure correctly, this won't really be a problem. Neither the agent nor you, for that matter, should have access to permanently delete the database and cause irreparable damage to any infrastructure. That's usually a sign of a poor choice of infrastructure architecture rather than a problem with the programmer or agent coding.
OpenClaw bots
Another use I have for Codex is that I use it for my OpenClaw bots. One of the great benefits of using Codex over Code Claude is that you can enable your OpenClaw bots with a Codex subscription, which you are no longer allowed to do with your Code Claude subscription. This is important because, in my opinion, Codex is a smart border-level model that you can use for your OpenClaw bots, which is also reasonably priced.
By this, I mean that the values of the Claude Code API do not work for almost all programmers out there, and therefore it is not an option for OpenClaw. Instead, you can buy a $100 or $200 subscription with Codex and have a very smart model to power your OpenClaw bots, which I believe is a good investment to make.
I also use fast mode on my OpenClaw bots since I have enough budget available to use it. However, you can also disable it if you think it is necessary, and, of course, it depends on the usage situation. In some cases, you rely heavily on quick responses from your coding agent, and in other cases, you just do a firework and forget it. The time it takes to do the work is not really important.
Work trees
Unfortunately, the OpenAI Codex hasn't implemented a simple job tree setup, like Claude Code did. This is definitely a no-go in my book, as task trees are an important feature to have when working on multiple items in the same repo at the same time.
However, to combat this problem, I set up a simple alias where I build my task tree when spinning Codex. I did this by asking Codex to create an alias for me so that when I type the command you see below, it will spin the task tree with the given name.
codex-wt This was very easy to set up and only took a few minutes with Codex.
Codex vs Claude Code
In this final section, I want to cover Codex vs. Claude Code and my opinion on comparing the two coding agents and frameworks. In my opinion, there is no clear winner between the two models. Both are extremely powerful, and in my opinion, I can complete even my most complex tasks with both models. However, I do have certain preferences in certain situations.
If I have a specific task I want to complete or I'm looking for specific bugs, in my opinion, Codex works best and is very efficient in completing the task. In most cases, Claude Code will also be able to complete the same task, but it just takes longer, from my experience.
In addition, as I mentioned in the OpenClaw section, the Codex allows you to use subscriptions to OpenClaw bots, which the Claude Code does not allow you to do. If you're leaning towards using more OpenClaw bots, I highly recommend doing it with Codex.
On the contrary, I think Claude Code is extremely powerful and can handle all my complex tasks while still having many features that I really enjoy. The job tree feature, for example, is a great addition from Claude Code, as well as the agent view, which they recently released. In my opinion, the Claude Code feature is more powerful than Codex and can be the reason to choose Claude Code over Codex.
However, all in all, I believe these two types are neck and neck and both are very strong. We'll have to follow both models in the future, keep testing them, and see which one comes out on top in a few months. For now, I believe both are good options, and what works best for you depends on your situation and preferences.
The conclusion
In this article, I discussed how to get the most out of the OpenAI Codex. I discussed why I started using OpenAI Codex, highlighting how I need to stay on top of the latest coding models and want to compare it to Claude Code. I got a very good first impression, I realized that the model was able to perform even the most complex tasks that I am working on. Then I put together some techniques I use to make the Codex better, such as:
- allowing it to verify its function
- to set the task tree alias
- i use it for my OpenClaw bots
Finally, I also had a section comparing Codex vs. Claude Code, where I highlighted that it is very close, and which is the better model. Which model is best for you depends on your preferences. I recommend that you look at both models to see what works best for you, and follow the models closely, because I believe that many new features and powerful LLMs will be available.
👋 Touch
👉 My Free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)
📚 Get my free ebook Vision Language Models
💻 My webinar on Vision Language Models
👉 Find me on social media:
💌 Stack
🐦 X / Twitter
