We notice the abnormal regression in behavior from the models

Codex
1

Over that last couple of days 5.5 xhigh has turned from a highly capable agent, to a toddler which can barely follow multistep instructions and implementation. Never ever have I experienced such utter frustration when trying to work with this model. The difference is extremely noticeable.

As I type this out, I have it running on the exact same task for the 10th time. The task itself comprises of 3 different things, which are not even moderate difficulty. Each time it has forgot 1 or even 2 of the 3 things. Other times it has been extremely lazy and just decided not to do one of the 3 things. This is not 5.5, it never did this. This behavior has been consistent for the last few days.

Openai should be sued for doing things like this. I don’t pay 200 a month for a functioning model to be degraded.

Reasons why this is happening:

  • nerfing the capabilities of 5.5 to make 5.6 seem better than it actually is.
  • cost reductions hoping that no one notices
2

i am noticing that it no longer/does not monitor subagents work proactively, it stops on minor issues instead of fixing them. it does not identify fixes but just stops with here are the issues.

3

Switch to GPT-5.5 medium.

4

Did you switch to higher abstraction? Be honest. Because I did.. for me it seemed logical that if model can do X it should be capably of Y where Y was a higher abstraction and not the same.

5

same experience and today I got the answer. looks like they are mixing 5.5 with 5.4 and 5.4 mini on every turn. the following shows my usage

image
image1402×554 29.6 KB