Why I swapped from Opus to Codex

Codex vs Opus is one of the most common debates currently.

I’ve used both Opus 4.5 and Codex 5.2 (and now Codex 5.3) extensively, both for over a month each, 12+ hours per day, so I wanted to share my thoughts.

First off is that they’re both really good models. Just because I use Codex now, it doesn’t mean Opus sucks. I also think you need both currently as they have different strengths, which I’ll get into shortly.

When to use Opus

Opus with the frontend design skill just destroys Codex at frontend, so I never use Codex for frontend tasks. It’s just way less creative, all of its UIs look the same, and it’s not as good at following existing codebase styling.

Every side project I do, I build with Codex, then pretty much run something like this with Opus:

“redesign the entire website with the frontend design skill, you don’t need to keep any existing designs/theme”

Another thing Opus is better at is just general computer use and agentic workflows outside of coding, but I can’t comment too much here as I pretty much only do coding with these models - this is just the consensus that I’ve seen online.

I pretty much don’t use Opus for anything else.

Downsides of Codex

Besides the previously mentioned frontend disadvantage, it’s dry to talk to which makes it hard to collaborate with sometimes.

Not only that, it’s worse at inferring intent so it sometimes says “want me to run these commands” as if it doesn’t know that it can do that already, whereas Opus would’ve easily ran them.

Why Codex is so good

When it was still Codex 5.2 vs Opus 4.5 I could somewhat see why a lot of people still used Opus 4.5. Codex 5.2 was really slow and barely showed what it was doing so you couldn’t see if it was going on some weird path.

Now with Codex 5.3, I think it’s pretty clear that it’s better than Opus 4.6 at coding. I didn’t spend too much time testing Opus 4.6 because Codex 5.3 was just way too good, and Opus 4.6 is just more token hungry than Opus 4.5, so the context window feels really small.

Here are the main advantages after testing both:

Context window/compacting/token efficiency. With Codex 5.3 I feel like I can keep a chat open for an insanely long time since it takes so long to fill the context window, and it can also survive multiple compactions just fine. With Opus I feel like 1 medium sized task and it’s about to compact.

Working style. From experience working with a large codebase as well as smaller ones, Codex is slower because it takes a lot of time to be thorough and ensure the task is completed correctly, whereas Opus feels like it’s rushing to get it done. This introduces way more bugs and actually leads to it being slower than Codex from experience.

Long running tasks. As mentioned before, Codex’s context window feels like it takes way longer to fill up as well as having a larger context window, so it’s just naturally better at longer tasks.

Usage limits. Codex limits are doubled for 2 months. It’s nearly impossible to run out currently. I’m running multiple chats in parallel and spamming multiple /review chats in parallel and I’m still struggling to run out.

Codex app. The Codex app was released recently and it’s really good despite some bugs. It’s the best agent orchestration tool right now and even lets you pick if a new chat is on local or a worktree, as well as having automations (cronjobs) that can work on different repos.

The reason why Codex 5.3 is so much better than 5.2 is because it’s now faster, but it also shows what it’s doing a lot more often, so you can stop it before it goes crazy. Not only that, it jumped from 64% to 77% in the Terminal Bench 2.0 benchmark. Benchmarks are usually useless but I can feel that it’s quite a bit smarter and it’s pretty much unmatched in intelligence outside of GPT Pro currently.

Overall, I always use Codex except for frontend because I know it’s more thorough and I can trust that it’ll get the job done. If my prompt is good, it’ll get it correct, simple as that, and it’s a breath of fresh air compared to Opus taking shortcuts and calling the task done.

Give both a try and see how you go, and feel free to message me if you need any advice.