A unfettered “o1” model framework, not released, and here running for many hours on 10000 candidate answers for six questions, could be of even more extreme operational cost, and thus a practical one is offered:
(IQ hit by safety training penalty after this is also possible)
The expanded availability of the o1-based models and more features is a hopeful possibility, addressed in the announcement. “I want” is tedious.
Without the ability for a developer to provide guidance, original behavior, and domain constraints, and its denial of a large domain of API use-case (and simply that it performs far better on a single benchmark-style input than on user language and iterative chat, in both quality and timely presentation) makes it an enigma to find application or value-add.
Like OpenAI’s contest process, I would put a user facing chatbot in front of the API model, to give input customized for it and not wastefully employ it outside of its reasoning specialization, like an auto-switching ChatGPT that OpenAI sees as a possibility.