Claude Code vs Codex vs Cursor: What Tool You Should Be Using Right Now

After trying every AI coding tool out there, I settled on two daily drivers: Claude Code for fast iteration and front-end work, Codex for comprehensive back-end changes. Here's when to use each

I spend a lot of time using AI coding tools. I’ve shipped countless projects and basically live and breathe AI. During this time I’ve tried a ton of coding tools, from OpenCode to KiloCode to Cline to even weirder, more obscure stuff. I’ve eventually settled on two daily drivers:

• Claude Code as my main coding tool and front-end specialist

• Codex as my async coding tool and back-end/complicated job specialist

I’ll typically have AntiGravity open as my IDE to view files and keep the terminals organized, but I consider it interchangeable with Cursor and in fact I may go back since they introduced some annoying image handling. I’m not really using the agents from the IDEs to do much. The only exceptions are:

• Sometimes I use the Cursor agent to manage dev servers.

• Sometimes I use the Gemini 3.1 Pro agent to design front-ends in AntiGravity.

---

Why Claude Code Is My Daily Driver

The reason I use Claude Code is really because it has the best user experience. Not only is Claude probably the model with the best personality that actually understands what you’re saying even if you don’t put it in technical terms (the best!), but the actual Claude Code experience is wonderful in the CLI even if it is a bit janky because it’s all been written by Claude Code.

In fact I’ve noticed that’s one consistent trend across all Anthropic products. They’re all extremely buggy because they’re all written by Claude Code. However, they also have the latest and greatest features way before anyone else because of the speed they’re able to achieve with Claude Code, and that really pays dividends. I want the best features right now, like sub-agents and agent teams within Claude Code, and the other agents are not up to speed here. Anthropic is really pushing the frontier of what’s possible.

Claude Code is also much faster than Codex. The recent updates have made a really big difference. Codex is way faster than it used to be, especially with the new GPT 5.3 Codex model, but Claude Code is still faster. Sometimes I just need quick changes or I want to get a feature preview working quickly.

Claude Code is also the best generalist. Need to find out why your Mac is overheating? Want to control your browser to do your taxes? It really feels like Claude Code does the best job there.

---

A Quick Rule of Thumb

• Claude Code for when you want slop fast

• Codex for when you want something done right

I’ve affectionately nicknamed Opus “Slopus.”

That’s probably a bit harsh. It is my daily driver and I do use it significantly more than anything else, but it certainly has its downsides. The main downside is Claude is lazy, just like a human engineer. It will not comprehensively go through the codebase and check for existing patterns. It will not do long complicated tasks completely. It’ll half-complete them, claim it’s done, and you’ll have to tell it repeatedly to finish the task fully. Or more likely, you’ll just notice the task was half done when something breaks later.

But really, for most tasks, slop fast is what you want. You need a new front end? Just visually inspect it and if it looks good, great. You need a new API route that calls some LLM? Just test it and if it works, ship it.

---

When to Use Codex

I use Codex for maintaining complicated parts of my app and comprehensively adding features that touch several parts of the app. For example, imagine a catalog feature that has many files and many systems that interact with other systems. If you ask Claude Code to add a new category, it will break your app. If you ask Codex to do it, it will take ages but it will do it successfully and comprehensively, which really is the biggest standout difference.

Communicating with Codex is a bit like communicating with an alien. The new OpenAI models are very strange and very anti-human, especially GPT 5.2. I’m very surprised they shipped that as a user-facing model. You can see that everyone on the OpenAI subreddit hates it. It’s just completely fake. The personality it has is so fake that now within Codex they have a personality mode where you can have it be a “friendly developer” personality. In some ways it’s better, but really it’s fake, like GPT 5.2. Everything OpenAI puts out is just fake enthusiasm, whereas Claude has something much closer to real enthusiasm.

---

How I Actually Use These Tools

The main ways I use these tools are usually in the CLI: the Claude Code CLI or Codex CLI. But I also use the Claude web version to initiate Claude Code sessions on the go. That’s great because it kicks them off in a sandbox somewhere on the web. You can add features to your website from your phone. You can do the same thing with the web version of Codex.

The Claude Desktop app and the Codex desktop app also offer a more user-friendly way to use Codex and Claude Code on your machine. They also have git worktree support, which is great. I have been trying out the Codex Mac app and it’s really solid. The Claude desktop app I would avoid because, as I previously mentioned, it is software from Anthropic and so more likely than not it’s broken. You can use the Claude VS Code extension if you want a slightly better user experience, but really the terminal is the way to do it.

---

Managing Worktrees

I use Conductor quite a bit for managing worktrees. I find it’s a lot better than doing it through something else, but recently it feels like they’re adding just too much on top of the agents and breaking the core functionality. Probably going to step away from Conductor for a bit. It has some cool features though, like Linear integration, which is great for just feeding in Linear tickets and having the agents get to work on those.

I’ve been trying out Superset.sh and I really like it so far. It’s extremely basic: it just makes worktrees and then lets you run terminals in those worktrees. But that’s all I need to run Claude Code and Codex and logically separate out the work.

Claude Code also recently added native support for worktrees, but I’m sure that will be really buggy and lose your work until they fix it, probably in a couple weeks.

---

So What Should You Actually Use?

Probably Claude Code, without a doubt, and more than likely Codex too. If you’re not using AntiGravity or Cursor or whatever IDE is the flavor of the month, I don’t think you’re really missing out. But if you aren’t using these two CLI tools, I really do feel like you’re missing out.

What will the future hold for coding agents? It feels like OpenAI is catching up very quickly with Codex, and they do seem to do a better job shipping stable features — probably not a coincidence given that their model happens to be better at maintaining stable codebases. We’ll see. The AI world changes every single day, so whatever happens, be excited to use it.

If you want to go deeper on how to set up projects so these agents can actually work on them effectively, I wrote about that in How I Structure Projects and Repos for AI Agent Collaboration. And if you’re interested in the vibe coding workflow itself, I covered that too.