OpenAI shows o1, an AI model that thinks before it answers

SAN FRANCISCO, SEPTEMBER 2024 - OpenAI, the company behind ChatGPT, has released a new kind of AI model. It is called o1-preview and is built to think first and only then answer. Subscribers to ChatGPT Plus and Team can use it from today. OpenAI does not call the model a successor to GPT-4o, the model that currently runs behind ChatGPT, but an addition to it for difficult questions.

How ChatGPT has answered until now

A language model such as ChatGPT writes its answer word by word. Each time, it picks the word that best fits the text so far. This is extremely fast, and for ordinary questions it works very well. With a task that needs several steps, such as a maths problem or a piece of a program, it sometimes goes wrong. The model starts to answer before it has the solution, and it cannot undo a mistake.

What o1 does differently

o1 first gets time to think. Before the answer begins, it writes out a long chain of thought, step by step, like a student working on scrap paper. Only after that does the answer follow. OpenAI says that during this chain of thought the model learns to recognise its own mistakes and to try a different approach when the first one does not work.

This ability comes from the way the model was trained. OpenAI used reinforcement learning for this: the model tries to solve problems and gets a reward when the result is correct. This is how it learns which way of reasoning leads to good answers. OpenAI does not show the chain of thought itself to the user. The company gives safety and competition as reasons. Some developers regret this, because they want to be able to check how the model reaches an answer.

Computing time as a new dial

OpenAI says that the accuracy of o1 rises when the model is allowed to take more time to think. Until now, a model mainly got better by making it bigger and training it with more data and more computing power. Now a second option is added: using more computing power at the moment the question is asked. Mira Murati of OpenAI calls this a new way of making models better.

What the figures say

According to OpenAI, o1 solves considerably more problems than GPT-4o. On the AIME, a US maths test that serves as a qualifier for the national maths olympiad, o1 scores 74 per cent of the points on average. GPT-4o reaches about 12 per cent. In programming competitions on the Codeforces platform, the model reaches the level of the best 11 per cent of participants. On questions about physics, chemistry and biology, it scores at the level of people with a doctorate, according to OpenAI. These are figures from the company itself, and independent measurements are still to come.

There are also limits. o1 is slower and uses more computing power than GPT-4o. OpenAI also released a smaller, cheaper version for maths and programming, o1-mini, which is about 80 per cent cheaper but has less general knowledge.

Looking back from 2026

On 5 December 2024, the full version of o1 appeared, together with a more expensive subscription with a more powerful version. In the same month, OpenAI showed its successor, o3. In January 2025, the Chinese company DeepSeek released R1, an open model that followed the same approach. In July 2025, a model from Google won gold at the International Mathematical Olympiad, a good deal harder than the AIME test on which o1 was still scoring.

What matters is the shift. A model does not have to become bigger to become better: it can also get better by thinking for longer. That makes AI better suited to work in which reasoning has to be done step by step, such as research and writing complex programs.