That's why they have two test sets. But OpenAI has legally committed to not training on data passed to the API. I don't believe OpenAI would burn their reputation and risk legal action just to cheat on ARC. And what they've reported is not implausible IMO.
