Reasoning models, and when they earn their cost

Two shapes of model on the same menu

Most products now offer at least two kinds of model, and the labels are unhelpful — "fast" and "thinking", or a name with "reasoning" or a "pro" suffix. The difference is real and worth understanding, because choosing wrongly costs either money or accuracy.

A standard model produces an answer directly. A reasoning model first produces a long internal working — often far longer than the answer you eventually see — and only then writes the reply. You are billed for that hidden working, and you wait for it. Answers that took two seconds take twenty, or two minutes.

The question is never which is better. It is which tasks pay for the extra tokens.

Where the extra thinking measurably helps

Tasks with interacting constraints, where an early choice must be revised once a later constraint is checked:

• A staff rota where six people have different availability, two cannot work together, and one must be on every Saturday.
• Working out whether a set of contractual conditions can all be satisfied at once.
• Arithmetic with several dependent steps — proration, apportionment, a tax calculation with thresholds.
• Finding a contradiction between page 4 and page 31 of a document.
• Anything where you would reach for paper.

The common feature: a solution can be checked more easily than found, and getting there requires holding several things at once. That is exactly what a long working is for.

Where it is money for nothing

Drafting, rewriting, summarising, translation, tone, formatting, classification with clear examples. Tasks with no search in them.

Pay five to ten times as much and wait a minute, and you get the same paragraph. Worse, on writing tasks a long deliberation sometimes produces a more laboured result — the model has talked itself into elaborateness.

A usable rule: if you would accept the answer without examining the working, you did not need a model that produces working.

The limit that surprises people

Extra reasoning does not manufacture missing knowledge.

If the model never saw the relevant regulation, no amount of deliberation will recover it. What you get instead is a longer, more structured, more confident wrong answer — the elaborate working makes the conclusion feel better supported while adding no information about the world. Some people find reasoning-model hallucinations harder to catch for precisely this reason.

Reasoning helps with thinking about material you supplied. It does not help with knowing things you did not supply. That distinction places most professional tasks correctly on the first attempt.

What it costs, concretely

Hidden reasoning tokens frequently run several times the length of the visible answer, and they are billed at output rates. A question whose answer is 300 tokens can consume 3,000 tokens of working. That turns a task costing a fraction of a penny into one costing a few pence — trivial once, and a genuine budget line when a team of forty routes everything through it by default because it sounds more thorough.

Some products expose a thinking budget or effort setting. Where it exists, use it: a medium setting on a moderately hard task is usually indistinguishable from maximum, at a fraction of the tokens.

A working policy

1. Default to the fast model.
2. Escalate when the task has more than about three interacting constraints, when it involves multi-step numbers, or when the fast model gave an answer you checked and found wrong.
3. Never escalate for tone, length or politeness. Those are brief problems, not capability problems.
4. When you escalate, check whether the answer actually changed. If the two models agree, you have learned something useful and you can stop escalating that task.

Point four is the one people skip, and it is where the savings are. Run the same twenty real tasks through both, compare, and you will usually find a small set where the reasoning model genuinely wins and a large set where it does not.

The free path

Open-weight reasoning models exist and run locally. The distilled DeepSeek-R1 variants and Qwen's reasoning models come in seven- and eight-billion-parameter sizes that run on an ordinary laptop through Ollama or LM Studio, both free.

They are slower on a laptop — expect to watch the working appear for a minute or two — and weaker than the best commercial reasoning models. They are also completely private, which for a rota containing staff names or a calculation over client figures may be the only acceptable option in your organisation. Slow and private beats fast and prohibited.