Following the long discussion as to whether GPT4 in ChatGPT is degraded. What seems to be occurring is that when you reached your limit previously you'd get a large notification saying that model has been switched.

They don't do that anymore and instead switch the model without notifying you or updating the default model notification at the start of the conversation.

So:

(a) it is degraded as it switches to 3.5-turbo

(b) it's evident in the speed of the inferences

(c) the quality drops as the context window is halved

(d) as you're notified of the change, it appears as if all is normal (as gpt4 is definitely faster than before), but in reality the instructions are followed less impressively and the context is lost faster

(e) this probably also accounts for the difference in experience across the comments as some people won't have hit the limit and been switched and some people will have switched.

No judgement on them doing this, but does explain the difference in experience.