A category of models has appeared in recent years: those that take time to work through a line of reasoning before answering. They solve problems the others miss, and they cost markedly more. Knowing when to use them has become a budget decision.
Definition
A reasoning model produces a series of intermediate steps before its final answer, instead of answering directly. Those steps are usually not shown to you in full, but they are produced, and they are billed.
The effect is clear on tasks requiring several linked deductions: a multi-stage calculation, a contradiction to spot across a document, a choice depending on several conditions.
The trade-off is easy to remember: these models answer better, more slowly and more expensively. On a simple task they add nothing, and you are billed for reasoning about a question that needed none.
When it is worth it
| Reasoning helps | Reasoning is pointless |
|---|---|
| Checking consistency across documents | Rewriting a text |
| A multi-stage calculation | Sorting into three categories |
| Debugging a failing scenario | Extracting an amount |
| Choosing between options | Translating |
| Planning an agent's steps | Summarising a short text |
The right-hand column covers most of a small business's uses. This is why reaching for the most capable model by reflex is a budget mistake: on these tasks a small model gives the same result for ten to fifty times less.
What to know before switching
Latency changes category. Several seconds instead of a fraction of one. On a Chatbot facing a waiting customer, it shows immediately.
Reasoning tokens are billed. They count as output, the expensive side. A short answer can cost a lot if the reasoning before it was long.
Displayed reasoning is not a justification. It is produced by the model like the answer itself, and it can contain a Hallucination. Convincing reasoning leading to a wrong conclusion is entirely possible, and harder to spot.
It is not a cure for factual errors. Thinking longer does not surface information the model does not have. For that you must supply it, through RAG or in the instruction.
On an automated task running thousands of times, measure before switching. The quality gap is sometimes nil while the billing gap is fivefold. That is exactly what an Evaluation on your own cases settles in an hour.
Frequently asked questions
How do you know a model is one of these?
Providers state it explicitly, and often let you tune the reasoning effort. On some recent models it is no longer a model choice but a setting, which simplifies the decision.
Should you use one for an agent?
For planning the steps, often yes: an agent choosing its actions badly stacks up pointless passes, and the reasoning pays for itself. For the actions themselves, a faster model is enough. See Agent loop.
Is the reasoning kept between calls?
No, it belongs to the current call. Like everything else it occupies the Context window while being produced, which matters on a long task.
How do you choose without wrecking the budget?
By starting with the cheapest model and moving up only when quality falls short, with measurement to back it. Our Claude Cowork course treats that choice as an arbitration to revisit regularly, not a final decision.