The AI tool wars have devolved into horoscope readings with syntax highlighting. People aren't comparing benchmarks. They're defending identities.
I have seen a similar developer commenting, "Claude is the best choice for React work" whereas another comment says, "Claude cannot be used for React, Codex outperforms it" Same person, same React framework, polar opposite impressions. Both are convinced. Neither has a controlled process.
The Unfalsifiability Problem
Here's what makes this astrology and not science: nobody controls for anything.
The outcomes you achieve using Claude are influenced by how you phrase your prompts, the complexity of your project, the programming language you use, the version of the library, and let’s be real – how you feel when you assess the results. And the same goes for Codex. The same goes for Gemini.
The performance of a model changes significantly depending on the language and framework used. For example, the experience of a developer using Rust differs from the one using Python. Additionally, the environment of a person developing a CLI tool is worlds apart from someone else connecting a Next.js app.
However, no one actually means "Claude is better for my specific use case with my specific prompting habits". What they claim is "Claude is better." End of story. 🙄
The Cult Dynamics
Both groups actually display very cult-like tendencies, if you've ever visited a developer forum thread about AI tools.
→ Cherry-picked examples presented as universal truth
→ Dismissal of contradictory experiences as "skill issue"
→ Identity fusion with the tool ("I'm a Claude developer")
→ Tribal hostility toward people who prefer the other option
This is not engineering. It's being a sports fan with a subscription fee.
The wildest part? People in the same thread, using the same model, report wildly contradictory experiences. One person says the model nails TypeScript generics. The next says it hallucinates types constantly. Neither is lying. Both are generalizing from a sample size of "my afternoon."
Your Workflow Is The Variable
Here's an uncomfortable truth nobody likes to hear: the model itself is not as important as the process you build around it.
It is more important how you break down problems before you cue. It is more important how you verify the output. It is more important if you provide context little by little or all at once.
I've had excellent outcomes with both Claude and Codex. I've had junk outcomes with both as well. But it was never about which model I used. It depended on whether I was lazy with my inputs or actually putting some effort and thinking into them.
The Real Question Nobody Asks
"What are the differences between these two models?"
→ What kind of code am I writing most often?
→ How do I structure my prompts — do I give examples or just describe?
→ Am I evaluating output carefully or just vibing on first impressions?
→ Could my "bad experience" have been a bad prompt? 🤔
These questions are dull. They don't spark engagement. They don't allow you to feel superior for choosing the correct team.
However, in reality, they can increase your productivity.
The Takeaway
Let’s not make model preference a personality trait. The way you prompt, the project you choose, and the way you evaluate the outputs are doing 80% of the work. The model is just the last mile. Pick whatever works for your context, revisit it regularly, and don’t dunk on the people who chose differently.
So what I'm asking is have you ever truly switched models for the same task with the same prompt and looked at the outputs right next to each other — or are you vibing as well? 👀
Top comments (1)
A Claude user here, agreeing with this post is not lost on me.
But that is kind of the point. I use Claude, I think it is genuinely good, and I still cannot tell you it is objectively better than Codex because I have never run a controlled test. I just know it works well for how I prompt and what I build.
That is a preference, not a verdict.