OpenAI releases o1, its first reasoning model

Until this point a large language model answered as fast as it could write. Quality came from scale: more parameters, more data, more computation spent during training. o1 was trained to do something else. Given a question, it worked through a long chain of reasoning that the user never saw, and only then wrote an answer.

OpenAI released o1-preview and o1-mini on 12 September 2024 to ChatGPT Plus and Team subscribers. The numbers it published were about hard problems rather than ordinary use. On a qualifying exam for the International Mathematical Olympiad, o1 scored 83% where GPT-4o had scored 13%. On competitive programming problems it placed in the 89th percentile.

The chain was hidden by design, so the work responsible for the improvement could not be read from outside. What could be observed was the trade: the model took longer and cost more per answer, and in exchange it was right more often on problems that reward being methodical.

The direction was not the one the field had been travelling. Since 2020 the reliable way to make a model better had been to make it bigger and train it on more text. o1 bought its gains with computation spent at the moment of answering. That opened a second axis, and it did not require training a larger model to move along it.

Trading time for accuracy had already been demonstrated that summer. DeepMind’s AlphaProof and AlphaGeometry 2 reached silver-medal standard at the July olympiad by searching for formal proofs, spending up to three days on a single problem where the human contestants had 4.5 hours. What o1 added was the same trade inside a general-purpose chatbot, sold by the month. Other laboratories shipped models that reasoned before answering within a few months of it.