o1: the year of reasoning models
Trained with reinforcement learning to think at length before answering.
12% → 83%AIME math: GPT-4o vs the full o1 (consensus over 64 samples)
Why it matters
Opened a new axis of scaling: not only more compute in training, but letting the model think longer at inference (test-time compute).
Costs and controversy
o1-preview shipped that day and the full o1 later; it is slow, expensive, and its thinking is hidden from users.
Built on
Led to
Source: openai.com ↗