In the International Olympiad in Informatics (IOI) competition, O1 was barely able to make progress on a problem that human competitors performed well on.

O1 demonstrated the ability to backtrack in its reasoning, where it recognizes a mistake and changes its approach, a capability not previously seen in autoregressive language models.

O1 performs worse than GPT-4 on certain tasks, such as creative writing.