On 12 September 2024, OpenAI released o1-preview and o1-mini, models trained to work through a problem step by step before giving an answer. OpenAI reported much stronger results on difficult maths, science and programming questions than earlier models.
What is different
Standard language models respond almost immediately. Reasoning models first produce an internal chain of thought, trying approaches and checking them, and then give a final answer. This takes longer and costs more computing power, but it can solve harder problems. Other companies have since introduced similar reasoning models.
Why this is useful for learning
- Showing the working. For maths and coding, the steps matter as much as the answer.
- Harder questions. Multi-step problems that stumped earlier models become more manageable.
What has not changed
Reasoning models can still be wrong. A longer explanation can sound convincing even when a step is mistaken. So the human still needs to understand the problem well enough to check.
What we teach
We ask learners to write down their plan before they write code. Breaking a problem into steps is the very skill that reasoning models imitate, and it is one that no tool can do for you if you want to truly understand.