GPT5.2 and your own perspectives ?

[lpalbou]

Hi there,

With a few colleagues, we were ecstatic about the performance of codex cli + 5.2xhigh, for us, the best cognitive architecture we have seen so far. Then, as it often happens in this rapidly evolving field, the performances degraded substantially.

To clarify : we are amongst those who think opus is a good model for toying and communication, but not for complex architecture and thinking. codex cli + 5.2xhigh was genuinely impressive and a generation or two above opus.

We had notice also that the 5.2-codex (and subsequent -codex versions) were really not on par with the 5.2 xhigh, but probably openai answer to "your models are too slow" (which is.. absurd, I would favor depth and accuracy to speed anytime, but that's another discussion). So I could definitely understand the drastic loss of performance if indeed instead of using 5.2 xhigh everywhere, they started routing to 5.2-codex...

My question to you : with your changes, do you somehow get back the performances you once had too ? or are there some other under the hood updates we can't access ?

And a second question maybe : I saw your branch for 5.4 with the same trick; which one do you prefer those days ? 5.4 is faster but "dumber" so far for me, at least with the regular codex cli.