OpenAI o1 — our first model trained with reinforcement learning to think hard about problems before answering. Extremely proud of the team! This is a new paradigm with vast opportunity. This is evident quantitatively (eg reasoning metrics are already a step function improved) and qualitatively (eg faithful chains of thought make models interpretable by letting you “read the model’s mind” in plain English). One way to think about this is that our models do System I thinking, while chains of thought unlock System II thinking. People have discovered a while ago that prompting the model to “think step by step” boosts performance. But training the model to do this, end to end with trial and error, is far more reliable and — as we’ve seem with games like Go or Dota — can generate extremely impressive results. It’s still early days for the o1 technology. It provides new safety opportunities which we are exploring actively, including on reliability, hallucinations, and robustness to adversarial attackers. For example, we’ve seen great uplift in our safety metrics by letting the model reason about policies via chain of thought. Its accuracy also has huge room for further improvement— for example, from our launch post, our model achieved 49th percentile / 213 points in this year’s competitive programming Olympiad (IOI) under human conditions of 50 submissions per problem. But with 10,000 submissions per problem, the model achieved a score of 362.14 — above the gold medal threshold. So the model is capable of even greater outputs than it appears at first glance.