Reported by @flat knot
(See screenshot in the replies below) The setup is in red, while observations are in yellow. The only way to reproduce this is to ask users who suspect they are intermittently encountering GPT-4o Mini (fast responses, no thinking, no tools) to repeatedly run a test. Choose a prompt which GPT-5 models would refuse but GPT-4o mini would comply (I'm not going to reveal mine since it is useful for holding companies accountable).
If it was GPT-5, it would have refused to reveal its system prompt. It would have reasoned, given that both GPT-5 Thinking + Extended Thinking were selected.
The model revealed its system prompt, revealing that it is in fact GPT-4o Mini. It did not think and wrote faster than 4o Mini.
Likely browser on Desktop (I got another user who was experiencing suspicious results to run it).