OpenAI o1-mini

What it is

o1-mini is a reasoning model that OpenAI released on 12 September 2024, at the same time as o1-preview^[1]. Both belong to the o1 series, and within that series o1-mini is the lighter option: OpenAI described it as smaller, faster and cheaper than its sibling^[1], and Microsoft put its cost at about 80% below o1-preview^[2].

The pitch was narrow and practical. o1-preview was the showcase model; o1-mini was the one meant to be cheap enough to call often, which is why the price gap mattered as much as the benchmark scores.

How it reasons

Like the rest of the o1 series, o1-mini is trained with large-scale reinforcement learning to reason using chain of thought^[3].

A chain of thought means the model works through intermediate steps before committing to an answer: it restates the problem, breaks it into parts, tries an approach and checks the result. Reinforcement learning is the training ingredient that produces this behaviour. Instead of only being shown examples of good reasoning, the model is rewarded for reasoning paths that end in correct answers, so it gradually learns which kinds of step tend to pay off. Much of that reasoning happens before the visible reply, which is why these models take longer to answer than a plain chatbot.

Under OpenAI's Preparedness Framework, the internal checklist used to assess risky capabilities before release, the model was rated medium risk overall^[3].

What it was good at

OpenAI described o1-mini as particularly effective at coding^[1], and the figures released with the launch support that emphasis. On a qualifying exam for the International Mathematics Olympiad, GPT-4o solved only 13% of problems while the reasoning model scored 83%^[6]. In Codeforces programming competitions, the new models' coding ability reached the 89th percentile^[6].

That mathematics result is the clearest illustration of what the o1 approach changed. The gap between 13% and 83% did not come from a larger model trained on more text; it came from a model allowed to spend time on the problem, testing steps before answering.

Access and limits

The model launched for ChatGPT Plus and Team users^[5]. After the 17 September 2024 update, ChatGPT limited o1-preview to 50 queries per week and o1-mini to 50 queries per day^[6] — a much larger allowance for the smaller model, consistent with its lower cost.

As an early model, the series lacked features such as browsing the web and uploading files and images^[6]. Users got reasoning, not a full assistant: no searching, no attaching a document, no looking at a picture.

Retirement

o1-mini had a short working life. OpenAI announced its deprecation on 28 April 2025 and shut it down in the API on 27 October 2025^[4].

Where it sits

o1-mini is best understood as the cheap end of the first o1 release, sitting next to o1-preview^[1][2]. It carried the same training recipe as the rest of the series^[3] into a package priced about 80% below the larger preview model^[2], and it took the coding workload OpenAI was most confident about^[1]. Its two-year-documented arc — launch, tight rate limits, a change of heart on availability, then deprecation — is a compact record of how quickly reasoning models were revised in this period.