The History of AI

OpenAI o1

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Alignment and reasoning

Lineage
Symbols and logic
Kind
Implementation

Who

OpenAI

What happened

OpenAI presents o1-preview and o1-mini, models trained with reinforcement learning to generate a chain of thought before answering. On the 2024 AIME math exam, o1 solves 74 percent of the problems with a single attempt, against 12 percent for GPT-4o. The full o1 arrives on December 5.

Why it mattered

The first commercial model trained to reason before answering. It opens a second route to better models: letting them think longer when they answer.

The full story

Until 2024 the main way to improve a language model was to train it bigger and on more data. Chain of thought showed that reasoning in writing helped, but models did it because the prompt asked them to, not because they had been trained to.

OpenAI trained o1, a project reportedly known internally first as Q* and later as Strawberry, with large-scale reinforcement learning so that it would use its chain of thought productively before giving an answer. According to the company, performance improved steadily with more reinforcement learning, that is, training compute, and with more time spent thinking at answer time, that is, inference compute. It was a new dimension along which to scale.

The results were a leap in mathematics and coding. On the 2024 AIME exam, o1 solved 74 percent of the problems with one attempt per problem, against 12 percent for GPT-4o, and 93 percent when choosing among a thousand attempts. In competitive programming it placed in the 89th percentile on Codeforces, and on a set of physics, biology, and chemistry questions it surpassed the accuracy of PhD-level experts.

OpenAI chose not to show the full chain of thought, only a summary generated by the model itself. In December it announced o3, its successor.

Videos

Builds on

  • Chain of thought (January 28, 2022)

    Inherited, with no proven direct link

    o1 is trained with reinforcement learning to make productive use of a chain of thought before answering

Leads to

  • DeepSeek-R1 (January 20, 2025)

    Inherited, with no proven direct link

    The same paradigm; no direct dependence is documented

  • Grok 3 (February 17, 2025), Claude 3.7 Sonnet (February 24, 2025), Gemini 2.5 Pro (March 25, 2025)

    Inherited, with no proven direct link

    After o1, the three labs add to their main models a mode that reasons before answering

Threads that pass through here

Sources

How to cite

“OpenAI o1.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/o1

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline