What is OpenAI o1, and what is a reasoning model?

On 12 September 2024 OpenAI released two new models, o1-preview and o1-mini, in ChatGPT and in its API. They are not the next number in the GPT line. OpenAI says this is a different kind of model, so it reset the counter and started a new series named o1. The idea behind the series is simple to state: before the model replies, it works on the problem privately, trying approaches, checking itself and backing up when a path fails. Only then does it write the answer you see.

That extra thinking costs time. A question that GPT-4o answers in two seconds can take o1-preview twenty seconds or more. In return, OpenAI reports large jumps on hard reasoning tasks. On a qualifying test for the International Math Olympiad, GPT-4o solved 13% of the problems; the new reasoning model solved 83%.

What "reasoning model" means

Researchers have known for a couple of years that language models do better on multi-step problems when you ask them to "think step by step." Writing out the intermediate steps, a practice called chain of thought, gives the model room to work. Simon Willison, a developer who writes widely about these tools, describes o1 as a specialized extension of that prompting habit.

The difference is that o1 does not wait for you to ask. OpenAI trained it with reinforcement learning, a method where the model is rewarded for reaching good outcomes, so that it learns to produce useful chains of thought on its own: breaking hard steps into smaller ones, noticing mistakes, switching strategy. OpenAI also reports that o1 gets better the longer it is allowed to think at answer time, not only the more it is trained. That second lever is what makes this a new category.

Where o1 shines, and where it doesn't

OpenAI's own numbers show the strengths clearly. On the 2024 AIME, a hard American high-school math exam, GPT-4o averaged 12% and o1 averaged 74% on a single attempt per problem. In coding contests on Codeforces, OpenAI places the model around the 89th percentile.

The same research post is honest about limits. When human reviewers compared answers side by side, they preferred o1-preview for data analysis, coding and math, but not for some writing and language tasks. At launch, o1-preview also cannot browse the web or accept uploaded files and images, and OpenAI says GPT-4o will remain the more capable choice for many common requests in the near term.

Here is a rough guide for choosing:

o1-mini is a smaller, faster version that OpenAI prices 80% below o1-preview. It is tuned for coding and reasoning rather than broad general knowledge.

The thinking you can't see

In ChatGPT you see a short summary of what o1 thought about, not the thinking itself. OpenAI explains that it hides the raw chain of thought for several reasons: it wants to be able to monitor that reasoning for safety, it wants to keep it free of the policy training applied to final answers, and it cites competitive advantage. Developers using the API are billed for these hidden "reasoning tokens" as output, even though they never see them, which makes costs harder to predict.

Access is limited for now. ChatGPT Plus and Team subscribers can pick either model from the model menu, with weekly caps of 30 messages (o1-preview) and 50 messages (o1-mini). In the API, only developers at the highest usage tier can try them, and features such as streaming, function calling and system messages are not yet supported.

Should you prompt a reasoning model differently?

Yes, a little. Many prompting habits from the GPT-4 era were ways of making the model slow down: "think step by step," "first list the facts, then reason, then answer," long worked examples. With o1, that thinking is already built in, so those instructions mostly repeat what the model does anyway. A clear, direct question with the facts it needs is usually the better prompt.

Keep the extra material lean, too. OpenAI's developer documentation, as quoted by Willison, advises including only the most relevant documents when you add background context, so the model does not overcomplicate its answer. That runs against the common habit of pasting in everything that might help.

What still matters is the same as ever: say exactly what you want back, give the numbers and constraints, and state how you'll judge a good answer. For example:

A bakery sells muffins at $3 each and coffee at $2. On Saturday it took in $410 and sold 40 more coffees than muffins. How many of each did it sell? Give the two numbers and one sentence checking that they add up to $410.

Notice what is missing: no role, no "think carefully," no step list. The request and the check are specific; the method is left to the model. If you need speed, files or a warm tone, switch back to GPT-4o. If you need a careful answer to a hard problem, give o1 a clean question and time to think.