Reasoning Models and Thinking
Level 4 · Choosing and running a model
Why some models write out their working before they answer, and what that costs. Which jobs it helps, and which it only slows down.
Course contentsLesson 12 of 15
Level 1 · What an LLM is
Level 2 · Inside the model
Level 3 · How a model is trained
Level 4 · Choosing and running a model
Level 5 · Where LLMs go wrong
After this lesson you can
- Explain what a reasoning model does first
- Set a thinking budget for a hard task
- Say when thinking is a waste
11 mincovers 4 glossary entries
Worth reading first: Post-Training and RLHF
Why some models think before they answer
A reasoning model thinks before it answers. It first writes a long chain of working, often thousands of tokens, then gives a short final reply. OpenAI's o1, in September 2024, made the idea famous. DeepSeek-R1, in January 2025, matched o1 on many maths and code tests, with open weights.
The thinking is ordinary text the model writes for itself. Some services show it, some show a summary, and some hide it. Either way you pay for it. Thinking tokens are billed as output tokens, the dear kind, even when you never see them.
Extended thinking is the switch some services give you for this. You switch it on and set a budget: a cap on how many tokens the model may spend thinking. A bigger budget gives better answers on hard problems, up to a point. It also gives a slower, dearer reply every time.
Behind it is an idea called test-time compute. Older models got smarter mostly through more training. Reasoning models also get smarter by doing more work while they answer. Give the same model more room to think, and it solves harder problems.
These models learned to think mostly through checkable rewards. They practised maths and code, and got a score only for a right final answer. They found on their own that long, careful working led to more right answers. Habits like checking work and trying again grew out of that.
Thinking pays on problems with many steps. That means maths, logic, planning, tricky code, and questions where the first answer that comes to mind is wrong. It wastes time and money on simple jobs. Ask a thinking model for a plain fact or a short rewrite, and it can take far longer for the same answer.
The written thinking is not a true record of how the model got its answer. Tests have found models leaving out a hint they used, or giving a reason that was not the real one. Read it as a useful draft, not as proof.
Worked example
Lucy turns thinking on for one step
Lucy's app plans delivery routes and also answers simple questions. With thinking on, its route plans pass her checks 9 times in 10, up from 6.
She keeps thinking off for the simple questions. There it only added cost and a longer wait, for the same answers.
The words for this lesson4
Check yourself
What do you pay for when a reasoning model thinks?
Finished this lesson?
Marking it complete also marks its 4 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.