What are reasoning models? Explained with a clerk's scrap paper
A reasoning model is an AI model that works through a problem in hidden steps before answering. Think of a clerk who works a sum out on scrap paper, then writes one clean figure on the form.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
This week's news that Google's Gemini 4 Hydrogen offers four "thinking levels" and that Meta's Muse Ember is a reasoning model with adjustable reasoning effort uses an idea now found in most new AI models, so here is what it means.
In one sentence: A reasoning model is an AI model that works through a problem in hidden steps before it gives its answer, which usually helps on hard, multi-step tasks but takes longer and costs more, because that hidden working is billed too.
The everyday picture
Think of an office clerk asked two questions. The first is "What is the date today?" The clerk just answers. The second is "How much do we owe this supplier after the discount, the late fee and the part refund?" Now the clerk reaches for scrap paper.
On the scrap paper, the clerk breaks the sum into steps, tries a figure, spots a mistake and works it again. Only then does the clerk write one clean figure on the form. You see the form, not the scrap paper. But the clerk's time was spent on both.
A reasoning model does the same. Its scrap paper is called thinking or reasoning, and it is written in tokens, the small chunks of text an AI model reads and writes.
You ask a question. The model works it out on hidden scrap paper (thinking tokens), then writes the answer you see. Both the working and the answer go on the bill.
How it works, step by step
- You send a prompt (the question handed to the clerk). With many models you can also choose how hard it should think, called reasoning effort or a thinking level (how much scrap paper to allow).
- The model reasons internally first (the working on scrap paper). Google's documentation says Gemini "reasons internally before responding." This working is made of thinking tokens, also called reasoning tokens.
- The raw thinking is usually kept back (the scrap paper stays on the clerk's desk). OpenAI says its reasoning tokens are not visible through its API. OpenAI, Google and Anthropic can return a short summary of the thinking instead.
- The model writes its final answer (the clean figure on the form). This is the part you read.
- You pay for both (the clerk's time on the scrap paper and the form). OpenAI bills reasoning tokens as output tokens, and Google says the price is the sum of output and thinking tokens.
Try it · Try the dial
How much scrap paper does this job need?
Drag the dial and watch what changes.
A task with a few connected steps. The default level is usually a sensible start. Test a lower one to save time and cost.
Where the picture stops
- A clerk thinks in ideas. A model's "thinking" is generated text in tokens. It often helps, but it is not proof that the final answer is right.
- A clerk can show you the scrap paper. Most AI providers return only a summary of the thinking, not every step.
- A clerk decides alone how much paper to use. With these models, you or the software often set the level, and the thinking counts toward the same output limit as the answer. OpenAI and Google both warn that if the limit runs out during thinking, the reply can come back incomplete.
Why it matters for your business
- Better results on hard work. The vendors point thinking at complex, multi-step jobs such as coding, maths, data analysis and planning.
- Cost you cannot see. Hidden thinking tokens appear on the bill. A higher level can cost more per request than the short answer suggests.
- Speed. More thinking means a longer wait. OpenAI suggests no reasoning for tasks that need speed, such as voice and classification.
- A dial to tune. Google advises low or minimal thinking for simple questions and its highest level for hard ones, so you only pay for depth where it helps.
An example
Imagine a 15-person property management firm using AI for two jobs. The first sorts incoming emails into "repair", "rent" or "other". The second checks a 40-page lease against the firm's standard terms and lists every difference. The firm runs the sorting at the lowest thinking level, because it is quick and the answers are simple. It runs the lease check at a high level, because it needs many careful steps. A person still reviews the lease list before anything is sent.
Try it · Myth or fact
Reasoning models: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 3 checked
Frequently asked questions
Is a reasoning model the same as a large language model?
A reasoning model is a kind of large language model. The difference is that it is built to work through hidden steps before answering, and many models let you turn that up, down or off.
Can I see what the model was thinking?
Usually only a summary. OpenAI, Google and Anthropic return summaries of the thinking rather than every raw step.
Do thinking tokens cost money?
Yes. OpenAI bills them as output tokens, Google adds them to the output price, and Anthropic counts them in the output tokens it bills.
Should I always use the highest thinking level?
No. For simple lookups or sorting, a low level is faster and cheaper. Save high levels for hard, multi-step tasks, and test on your own work to see where it pays off.
Related explainers
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.