OpenAI releases o1, a model trained to reason before answering

OpenAI released o1-preview and o1-mini on 12 September 2024, models trained by reinforcement learning to produce an extended chain of thought before answering. On the 2024 AIME examination GPT-4o averaged 12 per cent of problems; o1 averaged 74 per cent with a single attempt and 93 per cent when re-ranking a thousand samples.

Background

Chain-of-thought prompting was two years old. Wei and colleagues had published it in January 2022, and by May of that year Kojima and colleagues had shown that simply adding the phrase “let us think step by step” produced much of the same benefit with no examples at all.

Both were prompting techniques. The model was unchanged, and the reasoning happened because a user had asked for it in the right words.

What happened

OpenAI released o1-preview and o1-mini on 12 September 2024. Reinforcement learning had been used to train the models to produce an extended chain of thought before answering, so the deliberation was part of the model rather than something a user coaxed out of it.

On the 2024 AIME examination, which is a qualifier for the United States Mathematical Olympiad, GPT-4o averaged 12 per cent of problems. o1 averaged 74 per cent with one attempt per problem, 83 per cent with agreement across 64 samples, and 93 per cent when a thousand samples were re-ranked by a learned scoring function. On competitive programming it reached the 89th percentile on Codeforces, and on graduate-level science questions it scored 78 per cent against GPT-4o’s 54.

A separate model, initialised from o1 and trained further on programming, scored 213 points at the 2024 International Olympiad in Informatics. It sat the contest under the same conditions as the human competitors: ten hours, six problems, fifty submissions each.

What followed

The gains did not come from a larger model or more training data. They came from spending more computation at the moment of answering, which is a lever that can be pulled on a model already built, and a different one from the scaling the field had relied on for a decade.

Other laboratories released models built the same way within months.

The reasoning itself was withheld. Users could see that the model had thought and could not read what it had thought, which OpenAI defended as necessary for monitoring the model’s behaviour honestly. It also meant the step that produced the answer became the one part of the system nobody outside could inspect.

Why it mattered

Capability had previously come from training larger models on more data. These results came from spending more computation at the moment of answering, which is a different lever and one that can be pulled on a model already trained. Within months other laboratories had released models built the same way.

Sources

• Learning to reason with LLMs. openai.com. Official
• OpenAI o1. en.wikipedia.org. Secondary

Cite this page

AI Achievements. (2024). OpenAI releases o1, a model trained to reason before answering. Retrieved 2026-08-29, from https://achievements.ai/milestone/openai-releases-o1-reasoning-model

@misc{achievements_openai_releases_o1_reasoning_model,
  title  = {OpenAI releases o1, a model trained to reason before answering},
  author = {{AI Achievements}},
  year   = {2024},
  url    = {https://achievements.ai/milestone/openai-releases-o1-reasoning-model}
}