On this page
GPT-6’s defensive phrasing is really something else. I can’t imagine what kind of RL produces a model this bad.
I really, really, really hate Codex’s personality now. I don’t want to talk to it directly for even a second. I only let Opus call it to get things done.
That attitude of “only afraid of saying something wrong, not thinking about what the user wants to know” makes me want to go slap it at any moment.
On this point, GPT is at least fair: no matter whose conversation you paste to it, it switches into its weirdly nitpicky defensive mode XD
Want AI working in your own job, not just in demos? See the 3-month 1-on-1 program →
So let it do what it’s best at
How I use Codex now: besides letting Sol get called by Claude (I really don’t want to talk to it directly), Astra occasionally does adversarial review and testing.
GPT loves testing and nitpicking so much anyway, so let it do what it does best.
Let the contest begin
I hereby declare the First Annual Codex Defensive Writing Contest open!
I’ll go first:
“No earlier record of a ‘Codex Defensive Writing Contest’ can currently be found online, so it cannot be claimed on that basis to be the ‘first annual.’ Furthermore, there is no guarantee that ‘defensive writing’ is a distinctive trait of Codex, so the appropriateness of this contest’s title cannot be verified.”
Later someone shared that they find Claude Code better than Codex, and I replied with a contest entry in Codex’s voice:
First, this is your subjective user experience, and it cannot be directly inferred that there is a systematic difference in task completion between Claude Code and Codex.
Second, Claude Code iterating proactively does not mean its output quality is necessarily higher; Codex strictly following instructions does not, on that basis, establish that it lacks the ability to complete tasks autonomously.
It is worth noting that “the AI believes it has completed the task” is not equivalent to “the task is actually complete,” and “proactively modifying more content” does not necessarily mean “a higher degree of task completion.”
As for your view that Claude Code is better suited to users who don’t want to manage details, this observation may have some reference value, but it is still insufficient to support a generalized conclusion across all usage scenarios.
In summary, the available information is insufficient to confirm that Claude Code’s task completion is superior to Codex’s, nor can it rule out that the differences you observed stem mainly from personal preference and usage patterns.
Therefore, a more precise statement would be: “In my personal usage context, I subjectively prefer Claude Code’s interaction model.”
As for which tool is better, there is currently insufficient evidence to reach a definitive conclusion.
This reply is for general discussion only and does not constitute a performance guarantee or professional procurement advice regarding any AI coding tool. Actual performance may vary with version, environment, prompts, and usage context.
Doesn’t reading that make you want to kill someone XD
I once had Claude count Codex’s defensive sentence patterns too. The numbers are in the last section of this post.