After whatever quota of free GPT-5 messages is exhausted, `mini` should answer most replies, unless they're policy sensitive, which get full-fat `GPT-5 large` with the Efficient personality applied, regardless of user settings, and not indicated. I'm fairly confident that this routing choice, the text of Efficient [1], and the training of the June 2024 base model to the model spec [2] is the source of all the sophistic behavior you observe.
[1] <https://github.com/asgeirtj/system_prompts_leaks/blob/main/O...>
[2] <https://model-spec.openai.com/2025-02-12.html>