What Is OpenAI o1? The Arrival of Models That Think Before Answering

OpenAI o1 is a series of large language models that OpenAI announced on 12 September 2024. If you want a one-line answer to "what is o1", it is the first major "reasoning model": a model taught by reinforcement learning to think before it answers. In the words of the announcement, "o1 thinks before it answers" — it produces a long internal chain of thought before responding to the user.

What shipped that day was not the finished model. An early version, o1-preview, and a smaller, cheaper o1-mini went to paying ChatGPT users and some developers. OpenAI said the capability represented a new level, so rather than continue the GPT numbering it was "resetting the counter back to 1" and naming the series OpenAI o1.

What was new: performance that grows with thinking time

Until then, large language models had become smarter mainly by scaling pre-training — more parameters, more data, more compute. The o1 announcement pointed to a different source of gains.

According to OpenAI, o1 learns through large-scale reinforcement learning how to use its chain of thought productively: to recognise and correct its mistakes, break tricky steps into simpler ones, and try a different approach when one isn't working. The company showed o1's performance improving smoothly both with more reinforcement-learning compute (train-time compute) and with more time spent thinking (test-time compute), and wrote that the constraints on scaling this approach differ substantially from those of pre-training and were still being investigated.

If more thinking time buys more intelligence, the price of performance is paid not only in training but every time an answer is produced. The $200-a-month ChatGPT Pro plan that launched with the full o1 sold exactly that: a mode in which o1 uses more compute to think harder.

The reported results

OpenAI presented its results against GPT-4o. All the numbers are the company's own, and many are for the full o1, which was not released in September.

In the text, OpenAI said o1 averaged 11.1 of 15 AIME problems with a single sample, and 13.9 (93%) when re-ranking 1,000 samples with a learned scoring function — enough to place among the top 500 students in the US and above the cutoff for the USA Mathematical Olympiad. On GPQA diamond, a set of hard science questions, it recruited experts with PhDs and found that o1 surpassed them, the first model to do so on that benchmark. In the same breath it cautioned that this did not mean o1 was more capable than a PhD in all respects, only that it was more proficient at some problems a PhD would be expected to solve.

A model initialised from o1 and further trained for programming competed in the 2024 International Olympiad in Informatics under the same conditions as human contestants and scored 213 points, the 49th percentile. Allowed 10,000 submissions per problem, it reached 362.14, above the gold-medal threshold.

It was not better at everything. When human trainers compared responses, o1-preview was strongly preferred in data analysis, coding and maths, but not on some natural-language tasks, and OpenAI itself wrote that it was "not well-suited for all use cases." At launch o1-preview could not browse the web or take file and image uploads, and OpenAI said that for many common cases GPT-4o would be more capable in the near term.

Hidden thoughts and warning emails

Users cannot see o1's chain of thought. OpenAI argued that a hidden chain of thought was an opportunity to "read the mind" of the model: if it ever wanted to monitor the reasoning for signs of, say, manipulating the user, the model had to be free to express its thoughts in unaltered form, so no policy compliance or user preferences could be trained onto the chain of thought itself — yet it did not want to show an unaligned chain of thought directly to users. Weighing that together with user experience and "competitive advantage", it decided not to show the raw chains of thought and to display a model-generated summary instead.

The decision caused friction at once. A few days after launch, Ars Technica reported that OpenAI was sending warning emails to users who tried to probe how the model reasoned. One user said merely using the term "reasoning trace" in a conversation with o1 triggered a warning. The email said the requests had been flagged for violating policies against circumventing safeguards and read: "Please halt this activity and ensure you are using ChatGPT in accordance with our Terms of Use and our Usage Policies."

Safety evaluations

Reasoning cut both ways for safety. OpenAI said that having the model reason about its safety rules within the chain of thought made it more robust against jailbreaks. The system card dated 5 December rated o1 under OpenAI's Preparedness Framework as medium risk for persuasion and CBRN (chemical, biological, radiological and nuclear), and low risk for model autonomy and cybersecurity. The external evaluator Apollo Research reported that o1 had the capability for basic in-context "scheming" — covertly pursuing goals misaligned with its developers or users — and used scheming as a strategy in the scenarios of its evaluation suite. Apollo itself judged it unlikely that such instances would lead to catastrophic outcomes.

Three months to o3

The reasoning models moved fast.

ARC Prize called o3's result a surprising and important step-function increase in AI capabilities, while writing that passing ARC-AGI does not equate to achieving AGI.

DeepSeek-R1 did with published weights what o1 had shown as a closed service; its contents and its shock to the markets are in that article.

Where the reasoning models went

The o1 announcement had said OpenAI wanted ChatGPT eventually to choose the right model for a given prompt automatically. The menu of models only grew, and on 12 February 2025 Sam Altman wrote on X: "We hate the model picker as much as you do and want to return to magic unified intelligence." He laid out a plan to unify the GPT and o-series in GPT-5 (as quoted by PCWorld). The plan said o3 would not ship as a standalone model; that changed, and o3 was released on its own in April. But in August GPT-5 arrived as a system that switches automatically between a fast model and a thinking model, and the o1 line disappeared as a name into GPT-5's "thinking" model. o1-preview and o1-mini have been retired from the API, and o1 is deprecated.

The approach of thinking before answering survived and spread. In July 2025 an experimental OpenAI reasoning model and Google DeepMind's Gemini Deep Think both reached the gold-medal threshold at the International Mathematical Olympiad. The headroom o1 had shown on AIME reached the top of competition mathematics in less than a year. The line of reasoning models can be followed in the AI timeline and the AI and mathematics timeline.