Comment by Bjorkbat

It's easy to miss, but if you look closely at the first sentence of the announcement they mention that they used a version of o3 trained on a public dataset of ARC-AGI, so technically it doesn't belong on this list.