What Changed
A standard language model produces its answer token by token in a single pass. It has no mechanism to pause, check its work, or try a different approach.
Reasoning models generate an internal chain of thought first — exploring the problem, testing intermediate steps, sometimes backtracking — and then produce a final answer informed by that process.
The reasoning is generated as tokens like any other output, but it is working rather than answer.
Test-Time Compute
The insight behind these models is that quality can be improved by spending more computation at inference rather than only by training a larger model.
This is a meaningful shift. Previously, better output meant a bigger model trained at enormous cost. Test-time compute means the same model can produce better answers when allowed to think longer — and the amount of thinking can be dialled per query.
Most reasoning models expose an effort setting, letting you trade cost and latency against accuracy on a per-request basis.
Where They Help
Genuinely better on: mathematics, multi-step logical problems, competitive programming, scientific reasoning, complex debugging, and any task where an error early in the chain invalidates the result.
The common characteristic is that the problem has a verifiable correct answer reached through steps, and that checking intermediate work catches mistakes.
Where They Do Not
No meaningful advantage on: creative writing, summarisation, translation, simple factual retrieval, formatting and most conversational tasks.
For these, reasoning models cost several times more, respond several times slower, and produce output of similar quality. Using one as a general default is an expensive habit.
They also cannot reason their way to facts they do not have. Extended thinking about a question requiring current information produces elaborate, well-structured wrongness.
Practical Use
Route by task rather than standardising on one model. A fast standard model for the majority of requests, a reasoning model for the subset that genuinely needs it.
Where the model exposes its reasoning, read it when the answer surprises you — the chain of thought usually shows exactly where an incorrect assumption entered.
Deciding When the Extra Compute Is Worth Paying For
Reasoning models cost more and take longer, so the useful question is which tasks repay it.
They repay it when the task has a verifiable answer and multiple steps. Mathematics, code that must run, logic puzzles, planning with constraints, and analysis where an intermediate mistake invalidates the conclusion. The extra compute is spent checking and revising, and checking only helps when there is something to check against.
They do not repay it on tasks with no single right answer. Drafting, summarising, rewriting, tone work, brainstorming — here additional deliberation produces something more laboured rather than more correct, and occasionally worse, because the model reasons its way out of a good first instinct.
They do not repay it on simple retrieval or classification, where you are paying a premium for deliberation on a task that needs none.
The practical rule: if you cannot describe how you would check the answer, the extra compute is unlikely to be buying you accuracy. Route by task type rather than defaulting everything to the most capable option.
What the Visible Reasoning Is and Is Not
Many reasoning models expose a chain of intermediate steps, and it is easy to over-read.
It is not a guaranteed causal account of how the answer was produced. It is generated text, and a model can produce a plausible chain that does not correspond to the computation that determined its output. Treating it as an audit trail is a mistake — particularly in a regulated setting, where explainability means something specific and stricter.
It is still genuinely useful, for three practical purposes: spotting where an assumption was introduced, seeing which interpretation of an ambiguous question was taken, and finding the step where a long calculation went wrong.
Read it for those, and verify the answer independently.
Two operational notes. Latency changes the product: a response taking tens of seconds needs a different interface from one taking two, and designing for the fast case then switching models breaks the experience. And cost scales with deliberation, so a task routed to a reasoning model by default will be a great deal more expensive at volume than the headline token price suggests.