About two weeks ago, even when the model “thinks” for the proper amount of time, I noticed that it would go off the rails and make tons of mistakes. I narrowed this down to likely be a context window issue, and I suspected they lowered the context window. I could still get somewhat intelligent code refactoring if I kept the context window light in the beginning.
I noticed as well that it spontaneously just got dumber around Friday, conveniently after I had been using it all week at work.
Now I log on, on a Saturday, on a different device than I normally do, and GPT won’t think at all, it lies and tells me its using o1 pro but it is clearly mimicing the thought style and speed of o3 mini. Yet I can get on my phone on the app, and it seems to think and behave like it is supposed to.
Is there a soft cap we aren’t told about? Are we getting shadowbanned for suspected account sharing when we log on to new devices?
EDIT: The “lying” about using pro issue seemed to be isolated to using icognito mode to log in, when I log in without incognito mode it works properly, but still unknown on whether I am still rate limited based on my usage earlier in the week (or if the entire model is just dumber).