Opus 5 exploded my use of Codex, lol

the pairing works for me too, but i dont think its because codex is smarter. its that it isnt the author. i had opus write two things today and handed both to a separate agent whose only instruction was to try to kill them. it killed both, correctly. a model reviewing its own work mostly agrees with itself, becuase it shares the exact failure mode that produced the thing. one caveat on the 95% though, and im saying this as someone who got the number he asked for. a reviewer briefed to find problems will find problems. i told mine to default to refuting whenever it was unsure, which is what made it useful, but it also means my hit rate is measuring my instruction at least as much as its measuring opus. the signal id actually watch is the opposite one. when the checker starts agreeing with everything, thats usually not the code getting better, its the reviewer reasoning from the diff you handed it instead of from the thing itself.

I’ve been using both heavily for more than a year, and honestly, I don’t remember either of them consistently identifying meaningful gaps in the other’s work. Even if you repeatedly ask Opus or GPT to review its own implementation or plan, it will almost always find something to improve. They’re simply too biased toward validating their own approach, especially when the prompt is framed the wrong way.

That said, there’s one thing I’ve found extremely useful, when Opus or GPT gets stuck trying to fix something, the other model will often solve it immediately. Sharing that solution back with the original session almost always gets a “sorry, I should have thought of that earlier” response. 😏

Exactly how I use them together too, even made a video about it haha: https://youtu.be/hCOWH0RDB4s?si=pVJuekxHNclaNlP1

A nifty tool that calls the CLI and delegates to Codex: https://gist.github.com/nafetswirth/203206a4378957c3053fd95fc27886a5

Absolutely agree to this,

Codex have become so good now that i can trust it to do exactly what i would needed ai to do a year ago a true companion it have its downs but man it does its job so well,

the way i use it with all the mcps cloud services with gcloud read access its so powerful with its sidepanel browser use and computer use,

i have found so many bugs done a full automation end to end tests automated, have found errors an debugged many logical bugs

I definietely need to add codex to my ecosystem , bc Anthropic tokens are killing me.

Interesting pattern, pretty surprising actually. if Codex is better at code reviews, isn't it better at writing the code too? why use Opus at all then.

Hey, I’m using Codex to write the code and Opus/Fable for code review. Same story - they find something every single time. They make a great combo, no matter which one you choose for the coder role and which one for the reviewer role.