o1 and reasoning models

Chain-of-thought prompting had shown that models did better when they wrote out their reasoning. o1, previewed on 12 September 2024, was trained to do it. Using reinforcement learning on problems with checkable answers, mathematics, code, science, the model learned to produce long private chains of thought, to try approaches, notice errors and back up, before giving a reply. On a qualifying exam for the mathematical olympiad it solved 83 percent of problems where GPT-4o solved 13; on competitive programming it reached the 89th percentile of human entrants.

The company's chart showed accuracy rising smoothly with the compute spent at inference, on the same logarithmic axes as the 2020 scaling laws. Training compute had been the lever for a decade; now there was a second one, and a model could be made smarter by letting it think longer. The chain of thought was hidden from users, a choice OpenAI defended on safety and competitive grounds.

DeepSeek showed in January 2025 that the method could be reproduced cheaply and openly, and every laboratory shipped reasoning models within months. The gold-medal olympiad results of July 2025 came from their descendants.