OpenAI claims that their AI can get a gold medal on the International Mathematical Olympiad. The public models still do poorly even after spending hundreds of dollars in computing costs, but we've got a super secret scary internal model! No, you cannot see it, it lives in Canada, but we're gonna release it in a few months, along with GPT5 and Half-Life 3. The solutions are also written in an atrociously unreadable manner, which just shows how our model is so advanced and experimental, and definitely not to let a generous grader give a high score. (It would be real interesting if OpenAI had a tool that could rewrite something with better grammar, hmmm....) I definitely trust OpenAI's major announcements here, they haven't lied about anything involving math before and certainly wouldn't have every incentive in the world to continue lying!

It does feel a little unfortunate that some critics like Gary Marcus are somewhat taking OpenAI's claims at face value, when in my opinion, the entire problem is that nobody can independently verify any of their claims. If a tobacco company released a study about the effects of smoking on lung cancer and neglected to provide any experimental methodology, my main concern would not be the results of that study.

Edit: A really funny observation that I just thought of: in the OpenAI guy's thread, he talks about how former IMO medalists graded the solutions in message #6 (presumably to show that they were graded impartially), but then in message #11 he is proud to have many past IMO participants working at OpenAI. Hope nobody puts two and two together!