What is OpenAI o3-mini, and what does "reasoning effort" mean?

Yesterday, January 31, 2025, OpenAI released o3-mini, the newest member of its "o" series of reasoning models. OpenAI first showed it in December, alongside a larger, more capable model called o3. o3-mini is now live in ChatGPT and in OpenAI's API, the service developers use to build apps.

Two things make this release worth your attention even if you are not a programmer. First, free ChatGPT users can now try a reasoning model for the first time. Second, o3-mini introduces a simple idea that explains a lot about how these models work: you can ask them to think more or less before answering.

What a reasoning model does

Most chatbots start writing an answer right away, one word after another. A reasoning model first works through the problem in a hidden chain of steps, checking and revising its approach, and only then writes a reply. That takes longer, but it tends to help on tasks where one early mistake ruins the result, such as a math proof, a physics question or a bug in code.

OpenAI built o3-mini specifically for STEM: science, technology, engineering and math. According to OpenAI, at medium settings it matches the larger o1 model on math, coding and science tests, including the AIME competition-math exam and GPQA, a set of very hard graduate-level science questions, while answering faster.

How "reasoning effort" works

Developers can set o3-mini to one of three levels: low, medium or high reasoning effort. Think of it as telling a student how long to spend on scratch work before handing in the answer. Low effort is quick and cheap; high effort spends more time and computation for a better chance of getting a hard problem right.

Here is how OpenAI describes the trade-off, with TechCrunch's caveats added:

Setting	What OpenAI reports	Good for
Low	About as good as o1-mini on math	Quick checks, simple code, everyday questions
Medium	Matches o1 on math, coding and science; ChatGPT's default	Most homework-style and work problems
High	Beats o1 and o1-mini on AIME math, but only by 0.3 points over o1; still below o1 on GPQA	The hardest problems, when waiting a bit longer is fine

In ChatGPT, o3-mini runs at medium effort. Paying users can also pick o3-mini-high from the model menu when they want the deeper version.

Expert testers preferred o3-mini's answers over those of o1-mini, its predecessor, 56% of the time, and saw 39% fewer major errors on difficult real-world questions. On average it answered in 7.7 seconds, against 10.16 seconds for o1-mini.

Who gets it, and at what price

Plan	Access from January 31
Free	Choose "Reason" in the message box, or regenerate an answer
Plus and Team	Replaces o1-mini; limit rises from 50 to 150 messages a day
Pro	Unlimited o3-mini and o3-mini-high
Enterprise	Coming in about a week
API	Select developers first, in usage tiers 3–5

In ChatGPT, o3-mini can also search the web and give links to its sources, a feature OpenAI calls an early prototype.

For developers, o3-mini is OpenAI's first small reasoning model to support the features apps need in production: function calling (asking your app to run an action), Structured Outputs (replies in a fixed data format), developer messages and streaming. It can read up to 200,000 tokens and write up to 100,000 (a token is a chunk of text, often part of a word). It cannot look at images, so OpenAI recommends o1 for visual tasks. TechCrunch reports API prices of $4.40 per million output tokens and $0.55 per million cached input tokens, 63% cheaper than o1-mini.

Why it matters this week

The timing is hard to miss. In the past two weeks, the Chinese lab DeepSeek has drawn worldwide attention with R1, an openly available reasoning model with very low API prices. TechCrunch notes that o3-mini does not beat R1 on every test: it leads on AIME and on SWE-bench Verified, a software-engineering test, only at high effort, and trails R1 on GPQA at low effort. R1 is also still cheaper per token.

What o3-mini changes is access. Until now, reasoning models in ChatGPT were reserved for paying subscribers. Putting one in front of free users makes this style of AI, slower but more careful, part of everyday use. OpenAI also says o3-mini was trained with deliberative alignment, a method that teaches the model to think through written safety rules before answering.

Try it: Give o3-mini a short logic or math puzzle you can check yourself, then ask GPT-4o the same thing. Compare the time each takes and whether the answers agree. If they differ, ask each model to show its key steps.

How can free ChatGPT users try o3-mini today?

Open ChatGPT and look under the message box for the Reason button. Turn it on before you send your question, and ChatGPT will use o3-mini instead of its usual model. You can also regenerate an answer you already have to see the reasoning version.

Save it for questions that reward careful thinking. A good first prompt states the problem fully and asks for a checkable result:

Prompt

A train leaves at 9:40 and arrives at 13:15, with a 20-minute stop on the way.
How long was it actually moving? Give the answer in hours and minutes,
then show the two key steps so I can check them.

Short, clear problems like this play to o3-mini's strengths. For open-ended writing, casual chat or questions about images, the regular model is often the better choice. And as with any AI, check important answers, especially numbers, before you rely on them.