OpenAI was involved in funding FrontierMath, a leading AI math benchmark, but this connection was undisclosed until OpenAI's o3 model achieved a significant success on the test, solving 25.2% of complex problems—far surpassing previous results. Epoch AI, the benchmark's developer, admitted to mistakes in transparency, noting that more than 60 mathematicians involved were unaware of OpenAI's backing. These experts believed their work was exclusive to Epoch AI. Despite this, a verbal agreement ensured some test problems remained private to maintain independent verifications, while OpenAI was restricted from using the material for training. The situation has raised concerns about transparency in AI collaborations, highlighting the complexities and costs involved in AI benchmarking.