This was with GPT-4o (ancient now by AI standards). The reason OpenAI refuses to bring it back, despite all the demand, is that when the conversation gets long enough it starts doing weird things like this.

Subsequent models have almost entirely eliminated this kind of thing.

Even GPT-5.4 is ancient compared to intelligence of 5.6 although those are only 4 months apart. We're talking about at least twice the capabilities and that's scary to achieve in such a short timeframe.