I was a Cursor agent power user for months. I even wrote the blog post on Cursor tips that thousands of developers reference every single week. Then Claude Code came out and became my go-to. But already, my go-to has changed again.
I didn't want it to happen, but let me explain why it did. I now use Codex as my daily driver. Let me break down the agents, features, pricing, and user experience.
The Agents Are Converging
To be honest, all these products are converging. Cursor's latest agent is pretty similar to Claude Code's latest agent, which is pretty similar to Codex's agent.
Cursor set a lot of the foundation historically. Claude Code made significant improvements. Then Cursor started copying those improvements — like the to-do list and better diffing formats — and Codex started copying them too. Codex is so similar to Claude Code that I genuinely wonder if they trained off Claude Code's outputs.
A few differences I've noticed:
- Codex likes to reason for longer but has faster token-per-second output
- Claude Code spends less time reasoning but has slower token-per-second output
- Cursor with GPT-5 spends a long time reasoning — sometimes annoyingly so
- Cursor with Sonnet spends less time reasoning and more time outputting code, though the output is a bit slower (especially with Opus)
One thing I really like about Codex is model selection. The Codex model is noticeably better at knowing how long to reason for different types of tasks — over-reasoning for a basic task is genuinely annoying. You can also choose between low, medium, high, and minimal reasoning levels. I prefer these options over Claude Code's two-model setup. And with Cursor, the sheer number of options can feel overwhelming.
There's something to be said for using a tool built by the same company that trained the models. They know how to use it best, they work closely with those teams, and they give you the best price because there's no middleman taking a margin.
Overall, the agents are pretty comparable. If you prefer Cursor, Claude Code, or Codex — I respect that. My slight preference leans toward Codex or Claude Code, and I think it comes down to those companies knowing how to optimize their own models.
Pricing: Where Codex Pulls Ahead
Codex comes with standard ChatGPT plans. Claude Code comes with standard Claude plans. On the surface, pricing looks similar — a free tier, a ~$20/month plan, and high-end $100–$200 plans.
Here's the key thing: GPT-5 is a significantly more efficient model under the hood than Claude Sonnet, and especially Claude Opus. By most benchmarks and real-world anecdotes, these models are all pretty comparable in output quality. But GPT-5 costs anywhere from two-thirds to half of what Sonnet costs, and roughly a tenth of what Opus costs.
That means Codex can offer you more usage at a lower price.
From what I continue to see and hear, Codex is significantly more generous with usage limits. A lot more people will be fine on the $20/month Codex plan than on the $17/month Claude plan, where users seem to hit limits quickly. Even on the $100 and $200 Claude plans, some heavy users still hit limits. With Codex, I've almost never heard of someone hitting limits on the pro plans.
And these aren't just coding plans. You also get access to Claude chat or ChatGPT respectively. With ChatGPT, you get one of the best image generation models, video generation models, and generally more polished products — like the ChatGPT desktop app I use daily versus Claude's slower desktop app that's clearly an Electron web wrapper.
I will say Claude has better MCP integration with lots of connectors you can click to install. But day-to-day, I'm a ChatGPT user.
The number one complaint I see with coding agents is running out of credits. Codex has a very strong edge here.
User Experience: Pretty Similar, With One Annoyance
Codex and Claude Code have a pretty similar terminal experience. But one thing drives me crazy about Claude Code: the permission system.
The fact that I always launch it with --dangerously-skip-permissions is an unnecessary risk, but I hate how they don't save permission settings. Codex recognizes when it's in a git-tracked repository and is permissive by default — that's really nice.
Outside of that, the terminal UIs are fine. Claude Code's is a little better, but it doesn't make much of a difference.
Features: Claude Code Has More — But Do They Matter?
Claude Code has more features. You can create sub-agents. You can create hooks. There's extensive configuration. You won't find any of that in Codex.
But here's a secret. The Cursor team once asked me: "What features do we need to add to get you back from Claude Code?"
My feedback was this: I learned from Claude Code that I don't care about features at all. I just want the best agent I can ask for what I want and get it most reliably.
Cloud Code has cool features. Cursor does too. I use them sometimes, but I don't miss them when I don't have them. All I need is an agent and a good instructions file. I've got a whole other post on writing a good agents.md file — and that's really it.
Speaking of instructions, one thing that really annoys me about Claude Code: they don't support the agents.md standard — only CLAUDE.md. Tools like Cursor, Codex, and others all support agents.md. Claude refuses to. It's annoying to maintain a separate file for Claude when every other tool respects the standard.
The Killer Feature: GitHub Integration
Now let's talk about the one major reason I prefer Codex over any other tool: their GitHub integration.
I used Claude Code's GitHub app for a while and overall my team thought it was lacking. It didn't give helpful code reviews. It was verbose and pointless. It didn't find obvious bugs. You couldn't comment at Claude to fix anything. It kind of sucked.
Codex, on the other hand, has been amazing. Once you install the GitHub app, you can turn on auto code review for any repo. It actually finds legitimate, hard-to-spot bugs and comments on them inline. You can ask Codex to fix them. It works in the background, lets you know when it has a fix, you can view it and update your PR right there, then merge when everything looks good.
The most important part: when I build a feel for Codex in the terminal, I know what kind of prompts work well. I have my agents.md optimized in ways I know Codex handles well. When I ask Codex to do things from the GitHub UI, it's the same experience I get from the CLI — just running on a server. Same configuration, same behaviors, all of it.
The next best integration here is Cursor's BugBot. It also finds very good bugs with nice buttons to fix in web or fix in Cursor. You really can't go wrong with either. But my slight preference is Codex because I'm going straight to the model provider — I know I'm getting the best price, and we can give ChatGPT access to all our developers with near-unlimited Codex usage, PR comments, background agents, and PR reviews.
My Verdict
My personal winner right now is Codex. I use it daily and I like it a lot.
The real question is: what do you think? Have you used Codex, the Codex CLI, or the background agents and PR bot? Has it worked well for you too, or not so much?