The 2 frontier coding brokers proper now, by an extended shot, are Claude and Codex; nevertheless, I’ve observed important variations in when the 2 fashions are superior, and I’ve observed actual downsides to every of them.
I’ve thus applied a kind of inner classification system that I exploit each time I begin a brand new process, the place I determine whether or not I needs to be utilizing Claude Code or Codex to carry out the duty.
On this article, I need to focus on this inner classification system that I am utilizing, sharing my ideas on when which mannequin is superior, and aid you get probably the most out of those coding brokers by utilizing them on the duties they carry out finest at.
Why change coding agent primarily based on duties?
Round six months in the past, it was a fairly straightforward choice on which coding agent you ought to be utilizing. Anthropic, with their Opus collection mannequin, was simply far superior in all coding duties. After all, you would be utilizing different fashions similar to Google’s mannequin for coding or Codex even at the moment, however for my part there was a really important distinction within the efficiency of those fashions in comparison with Opus.
Nonetheless, the aggressive panorama of coding brokers has modified considerably in these six months, and I now imagine that there are two frontier fashions, Codex and Claude Code, with a number of different attention-grabbing rivals very shut by. For instance, GLM 5.3 or Kimi K3, that are each wonderful coding brokers, although not fairly on the efficiency of Claude Code or Codex.
Thus I will be specializing in the 2 frontier coding fashions for now, although I imagine in a couple of months this may change and we would have one other frontier coding agent. Nonetheless, the learnings I will focus on on this article are fairly generic, and it is about the way to acknowledge when a coding agent performs higher and by which conditions a coding agent struggles. I will focus on the completely different weaknesses and strengths of Claude Code and Codex and likewise focus on how one can uncover these points and aid you select the perfect coding agent for the duty that you just’re engaged on each now and sooner or later, as soon as the panorama of coding brokers adjustments considerably.
Strengths and weaknesses of Claude Code and Codex
First, let’s focus on the strengths and weaknesses of Claude Code and Codex. To maintain it tremendous easy, I’d clarify the completely different conditions the place it is best to use every mannequin with the next sentence.
Codex is much superior when engaged on a single particular tough process that you just need to drive to completion, whereas Claude Code is superior at orchestrating brokers to shortly get by way of a bunch of smaller duties.
Now let me elaborate a bit on every level. I will begin with Codex. Total I feel I’ve a desire for the Codex coding agent at present, which is predicated on a couple of components: one is that Claude Opus 5 is manner too talkative, and I’ve to inform the mannequin to be extra concise in its responses a number of occasions per day, despite the fact that I’ve very robust factors in my markdown information highlighting that the mannequin needs to be concise.
I haven’t got this difficulty in any respect with Codex. Codex is extra straight to the purpose, and I additionally really feel a bent that Codex is extra keen to only get work accomplished as a substitute of asking me questions on a regular basis, whereas Claude leans extra towards asking me questions and stopping with out ending the entire work. No less than if I do not actively use the /aim command.
So basically, when driving a single, sometimes tougher process, I’ve a robust desire for utilizing Codex as a result of it simply has a greater capacity to get that stuff accomplished appropriately.
You may suppose that Codex having this trait makes it the superior mannequin in all coding duties. Nonetheless, sadly, I discover that now I am doing so many duties in parallel as a result of a variety of duties that are available in, sometimes by way of product suggestions, are smaller fast fixes that you do not want an excellent good mannequin to finish.
Naturally, I do not need to should manually spin up separate brokers for every such smaller process as a result of I can have between 50 and 100 such duties are available in every day, and it will take a variety of effort from me personally to spin up all of these classes myself.
Thus, I do wanna have an orchestrator agent that orchestrates sub-agents to finish every of those smaller duties individually. And that is the place I discover Codex actually struggles.
Codex is impressively unhealthy at orchestrating a variety of completely different brokers to get a variety of completely different smaller duties accomplished. On the whole, should you simply ask Codex to finish two duties, particularly if they don’t seem to be very strongly associated to one another, I discover that Codex many occasions forgets about one of many duties and would not full it.
This, in fact, makes Codex a hopeless mannequin with regards to organizing a variety of smaller duties and getting such duties accomplished. Thus, my high-level classification system works like the next.
For every day I get a variety of smaller duties in and I’ve a single Claude Code session the place I manage all these smaller duties and have Claude full them with sub-agents. Then each time I’ve larger duties coming in or larger initiatives, I at all times spin up a single Codex session per such undertaking or process and have that accomplished. Additionally I’ve a desire for utilizing Claude Code with regards to design duties or implementing entrance finish solely adjustments (although these are virtually at all times fast fixes, so I do them with Claude in any case)
Now I do need to notice that this may change very quickly. OpenAI may include some upgrades to their harness, or they may launch a brand new mannequin that’s stronger at orchestrating duties. And on this occasion, if so, I will transfer over to Codex full time, mainly.
Find out how to uncover the place a mannequin excels and the place it struggles
Now that I’ve mentioned my preferences on Claude Code and Codex and when to make use of every mannequin, I need to transfer on to a extra basic matter, which is the way to uncover the place a mannequin excels and the place it struggles. To begin off, I will spotlight how I found the problems I discussed above with each Claude Code and Codex.
On a excessive stage, I feel this matter is about typically being attentive to how your coding brokers work and, once they do work, analyzing what they do, how they did it, and the way lengthy they took. To do that evaluation, you’ll be able to, in fact, use a coding agent to, for instance, look into metrics similar to:
-
Common time to dev for a single process
-
Variety of PR assessment rounds
And plenty of different metrics, in fact. On the whole, you may also simply observe your instinct and see once you really feel like a process is taking longer than it ought to. For instance, one robust factor I observed is that after I was utilizing Claude Code to repair single duties, it had a robust tendency to at all times cease and ask me for stuff, despite the fact that I did not need it to. After which each time it requested me stuff, it included manner too many phrases, and it made it very tough for me to know what the mannequin really wished from me.
Thus, I began testing Codex on the very same duties and observed a stark distinction. It was extra in a position to simply full the duty and make assumptions that have been, for probably the most half, proper, which mainly made it more practical at finishing the duty for me. So, basically, what I did to check them is that I simply ran the identical process with each fashions, which, in fact, prices some additional tokens, nevertheless it’s price it to search out the optimum mannequin for a process that you just’re engaged on, at the least as one thing you are able to do from time to time.
And now, on the opposite facet, the way in which I found that Codex was unhealthy at orchestrating smaller duties and dealing on a variety of smaller duties was that I’d, in some situations, have Codex work on a single process, then I would ask for a small modification to that process or to repair one thing form of associated to that process, however on the facet, and I’d discover that on a surprisingly frequent foundation, Codex would simply merely neglect about doing one of many duties, and I would should remind it about it. And in lots of circumstances, it forgot the duty once more. After all, that is hopeless and typically tough to detect as a result of after I hand duties off to brokers, I anticipate them to recollect the duty, full it, and ask me for approval earlier than forgetting about it themselves.
Thus, I began orchestrating such duties with Claude Code as a substitute and testing the very same duties, and I observed it was significantly better at remembering excellent work that it needed to do and at orchestrating sub-agents to do a variety of smaller duties.
Conclusion
On this article, I mentioned when to make use of Claude Code and when to make use of Codex on your coding. At the moment, these are two frontier fashions for my part, although this may change considerably within the coming months, and particularly thrilling is that we’ve got a variety of open-source fashions that are performing extremely properly and at a a lot cheaper price level than frontier fashions. I mentioned the professionals and cons of each Claude Code and Codex, and after I use every mannequin with my inner classification system. I then began speaking a bit extra on the whole about how one can uncover the place a mannequin excels and the place it struggles. All of it comes right down to having a sense for when the fashions are performing properly and once they’re being sluggish. Moreover, I exploit extra quantitative measures by having a coding agent from time to time undergo my metrics, similar to common time to dev or what number of assessment rounds to get code to dev. I imagine it is best to run these checks frequently to guarantee that your tech stack is optimized.
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day E-mail Course)
📚 Get my free Imaginative and prescient Language Fashions e book
💻 My webinar on Imaginative and prescient Language Fashions
👉 Discover me on socials:
💌 Substack
