The old traditional way of coding (and learning to do it) is long gone. AI-assisted coding, or maybe vibration coding as it is now affectionately called, has become the default for a growing number of developers. Ultimately, the tools you choose are more important than ever, and copying code snippets from ChatGPT, Claude, and Gemini and then pasting them into your IDE is no longer enough. There is a plenty of dedicated AI coding tools nowand the most capable ones live directly in your terminal or editor. These tools can understand your entire code base and are capable of executing complex changes to multiple files from a single message.
Claude Code has quickly become my go-to tool for this purpose and has led me to spend more time in the terminal than Instagram (and I say this as someone who was once terrified of the terminal). Claude Code’s most direct competitor is OpenAI’s Codex, which is relatively new compared to Claude Code’s several-month lead. I tried it during its early days and it didn’t really match Claude Code in any meaningful way. Codex has improved rapidly since then, so I decided to ditch Claude Code for a week and go all-in on Codex to see if it could hold up as my daily driver. That’s how it was.
Do you want to stay up to date with the latest in AI? The XDA AI Insider newsletter is published weekly with deep dives, tool recommendations, and practical coverage you won’t find anywhere else on the site. Subscribe by modifying your newsletter preferences!
To start, Codex is open to everyone.
Try it before you buy
The main reason people around me haven’t used Claude Code is the barrier to entry. You need a paid account to try it out and see if it meets your needs, and not everyone is willing to sign up for a plan just to try a single feature (without knowing if it’s worth it). The fact that Claude does not currently offer a free trial option makes this even harder to justify. The only way to try it for free is if someone provides you with a referral link!
Codex, on the other hand, is available in the free tier of OpenAI. You can log in and start using it. If you like it and think it’s valuable to you, you can upgrade to a paid plan to use it more. If not, you haven’t lost anything. It’s a much easier ask than committing to an upfront subscription just to see what all the fuss is about. Now, to make this test fair, I also used a paid Codex level. But the fact that there is a free option gives Codex a huge advantage when it comes to getting people through the door. Claude Code could be the best tool in every other way, and it still wouldn’t matter if people never try it.
Codex is more autonomous and I’m not a fan
Slow down, I didn’t say build yet
I’m not a fan of giving the AI a vague indication and just… building what it thinks I want to build. I prefer to spend some extra time explaining my preferences, answering any questions, and simply giving as much detail as possible to ensure the result you produce is as close to my vision as possible.
In my testing, I found that Claude Code does this naturally. He asks a ton of questions, checks assumptions, and keeps you informed at all times. When he doesn’t know something, he doesn’t try to fill in the gaps himself and almost always asks for preferences before trying. To be clear, I’m not even referring to Claude Code’s Plan mode here. Instead, I’m referring to its default behavior. Even without explicitly telling you to plan first, Claude Code simply pauses and communicates with you.
I noticed that Codex, on the other hand, just takes your cue and starts building. For example, I recently created a tool for my workflow using Claude Code and decided to use the same prompt to see what Codex would produce. Claude Code asked about 10 questions trying to identify exactly what I wanted to build and how I wanted each aspect of it, while Codex started building after I sent the same message. The result was just as you expected. Claude Code’s result was significantly closer to what I actually had in mind, because I had spent that extra time understanding what I wanted before writing a single line of code.
On the other hand, while the Codex version pretty well encapsulated exactly what I wanted to build, it was also full of assumptions that I didn’t agree with. For example, the tool I built relies heavily on an AI model’s API to process inputs and return results. It’s a fundamental part of how everything works. Codex assumed I wanted to use an anthropic model for this, which honestly made me laugh. Is Codex, an OpenAI product, by default Anthropic?
Meanwhile, Claude Code took a completely different approach. He explained my options, explained the pros and cons between the different providers, broke down how much each would cost per call, and allowed me to make the final decision. I ended up choosing a configuration that made a lot more sense for my use case. Out of curiosity, I decided to add the same API key to the Codex version for comparison, and the version it produced unfortunately cost me more per task than the one Claude Code helped me build. This was because Claude had actually optimized for efficiency after understanding what he needed, while Codex simply chose a model and moved on.
me recently created the same app with Claude Code and Codex and wrote about the experienceand I noticed the same pattern. Claude Code’s willingness to ask questions from the beginning consistently led to better first drafts, while Codex’s rush to build meant more time spent fixing things after the fact.
Codex limits are much more generous
You can really finish what you started.
The biggest complaint you’ll hear from Claude users is the brutal limits. Their limits were always relatively strict, but they have gotten noticeably worse in recent months. There are a number of factors contributing to this, including the increase in users due to the Pentagon x OpenAI agreement and Anthropic refuses to comply with the War Department’s stance, Claude’s stance on AI safety, the creativity of the model, and the new features and capabilities it is constantly updated with.
While it is unfortunate for Claude users, the results of this surge have reached them and, unfortunately, that has resulted in even stricter limits. In fact, Thariq Shihipar shared a post on X Claiming that around 7% of users will now reach limits they didn’t have before! Despite being on the Max 5x tier which has, well, 5x the usage of the Pro plan per session, I still find myself hitting limits too often.
Here’s the interesting thing: In my testing, I found that Codex’s Pro tier limits were more generous than Claude Code’s Max 5x tier. Frankly, Claude Code’s Pro plan also has disappointing limits. In Claude Code, I hit the session limit with just a few prompts (even when using a lightweight model like Sonnet). That says a lot.
As for money, Claude has four tiers: Free, Pro ($20/month), Max 5x ($100/month), and Max 20x ($200/month). Codex also has four tiers that directly concern the average user: Free, Go ($8/month), Plus ($20/month), and Pro ($100/month and $200/month plans). So with Codex, you will definitely get more value for your money.
Claude Code comes with a complete ecosystem
You’re paying for much more than just code
At the rate at which Claude is developing, it is now starting to feel like a full-fledged ecosystem. With a paid subscription, you also have access to Coworkwhich are essentially the agent capabilities of Claude Code but for non-coding tasks. You can read and write to your local files, organize your folders, write reports from source documents, and even schedule recurring tasks like pulling metrics or running a weekly summary, all without you having to sit down and manage every step.
They Recently added officewhich allows you to send a task to Claude from your phone and have him do the work on your computer while you’re away. On top of that, there’s a plugin marketplace, MCP connectors for tools like Slack, Google Drive, and Microsoft 365, and even a brand new product called Claude Design for creating images like mockups and slides.
In other words, when you subscribe to Claude, you get a coding tool, a desktop assistant, a design tool, scheduled automations, and a growing library of integrations, all under one roof. The Codex doesn’t really have an equivalent to this. It’s a great coding agent and the GitHub integration is best in class, but that’s where it ends. If you’re someone who wants one platform to handle as much of your workflow as possible, Claude’s ecosystem is hard to match.
Codex could be the best option for experienced developers
Beginners, you might want to stick with Claude.
This point goes back to autonomy point I above. I think Codex’s approach works well if you’re an experienced developer who already clearly knows what you want and doesn’t need hand-holding to make decisions.
However, if you’re new to coding, that same approach can leave you with code you don’t really understand and decisions you didn’t know were being made. Claude Code’s habit of asking questions, explaining trade-offs, and guiding you through options is a genuinely educational experience and helps beginners like me immensely.
Looks like I’ll be renewing my subscription to Claude Code
Despite the better limits you get with Codex and the fact that anyone can try it for free, Claude Code is still the tool I would recommend to most people. That being said, I think Codex is great and worth trying out too. It’s just not the tool for me!
