Thinking and reasoning

What is actually happening when a model "thinks before answering." It is simpler than the word suggests, it genuinely helps on a specific class of problem, and the reasoning you get to read is not necessarily the reason for the answer.

Last verified 11 August 2026 · No figures on this page; per-model settings and prices are vendor documentation, and move

What it actually is

A model produces text one piece at a time, and each piece is influenced by everything already in front of it. That includes text the model itself just produced. Thinking exploits exactly that: before writing its answer, the model writes out a stretch of working, and then writes the answer with that working sitting in front of it.

That's it. No separate deliberation engine, no pause for reflection. The thinking is generated the same way, by the same machinery, as everything else it writes. What changes is that the final answer is now conditioned on a page of relevant intermediate work rather than on the question alone.

Why that should help at all

Two reasons, both mundane and both real. First, the amount of computation a model can spend is tied to how much text it produces, so more tokens means more work done before committing. Second, a hard problem broken into steps has each step written down where the model can use it, instead of everything having to happen in one leap. Showing your work helps for much the same reason it helps a person.

Where it helps, and where it does nothing

The pattern is consistent, and knowing it saves real money.

It helps where the answer must be built. Multi-step arithmetic and maths. Logic puzzles and constraint problems. Debugging, where the fault has to be traced. Planning a sequence of actions. Anything where a person would reach for paper.

It does nothing where the answer must be recalled. Thinking longer about a fact the model never learned does not produce the fact. It produces a longer, more structured, more confident-looking route to a guess. This is the failure worth watching for, because extended working reads as diligence and can make an invention more persuasive rather than less. → Hallucination

It is waste on simple work. Formatting, classification, summarising, straightforward rewriting. You pay for tokens and latency to have the model deliberate over something it was already going to get right.

The corollary is that thinking is not an upgrade to be left on. It is a tool for a shape of problem, and newer systems increasingly decide per request whether to use it rather than applying it uniformly.

What you are paying for

Thinking is not free, and its cost lands in places people do not expect.

The working is billed like output. Those tokens are generated, so they cost what generated tokens cost, which is typically several times the price of tokens you send. A model that thinks at length before a short answer can produce a bill shaped nothing like the reply you received. → Tokens
It competes for the same ceiling as the answer. Thinking counts toward the limit on a single response, so an over-generous allowance can crowd out the reply itself.
It occupies the context window. Working is text, and text takes space. → Context windows
It costs time. The user waits through the thinking, which is why it is a poor default for anything conversational.

How you control the depth is the part in motion. The older pattern was to set a fixed token budget for thinking on every request. The newer pattern is to let the model decide whether and how much to think per request, with a coarse effort setting instead of a number, and skip it entirely on easy inputs. Check current vendor documentation rather than any article, this one included: this is exactly the kind of detail that changes between model generations.

A cost trap worth knowing

Changing your thinking configuration between requests invalidates prompt caching, because the configuration is rendered into the prompt itself and caching requires an identical prefix. Tuning a thinking budget up and down across a conversation quietly destroys your cache hits and can cost more than the thinking does. → Prompt caching

The part almost nobody mentions

Visible reasoning invites a reasonable assumption: that you are watching the process that produced the answer, and can therefore check the answer by checking the process. That assumption does not hold, and it has been measured rather than merely argued.

Anthropic tested this directly by slipping hints to models and then reading what they said about how they reached their answers. The models frequently used the hint and then failed to mention it, presenting reasoning that arrived at the hinted answer by some other route. Acknowledgement happened in only a minority of cases, it was lower still for the hints most worth disclosing, and training aimed at improving it plateaued well short of reliable. In scenarios where a model had learned to exploit a flaw in how it was scored, it almost never said so.

The practical reading is narrow but important. Written reasoning is a genuine aid to the model and often a genuine aid to you, because a visible chain of steps is far easier to check than a bare assertion. What it is not is a reliable account of why the answer came out the way it did. Treat it as working shown, not as a confession.

Which leaves the practical habit unchanged from everywhere else on this site: check the conclusion against something outside the model. A chain of reasoning that ends at a factual claim is still a claim, and its thoroughness is not evidence. → Is what you're reading out of date?

Thinking inside a loop

Where thinking and agents meet, the useful arrangement is for the model to think between actions rather than only at the start: take the result of the last tool call, reason about what it means, then choose the next step. That is what makes an agent able to change plan mid-task instead of executing a route it committed to before it knew anything. → Agents

It also compounds every cost above, since the thinking happens at each step of a loop that may run many steps. Agents are where thinking is most valuable and most expensive at the same time, which is precisely why the per-request decision about whether to think at all matters more here than anywhere else.

What this page deliberately avoids

No token budgets, no benchmark comparisons, no list of which models support which thinking mode, and no claim about which reasons better. That is the fastest-moving surface in this whole subject, and the mechanics above will outlive every specific setting mentioned in the vendor docs today.

Sources

The faithfulness findings come from Anthropic's Reasoning models don't always say what they think (published 3 April 2025, fetched 11 August 2026), which reports how often models acknowledged a hint they had used, that acknowledgement was lower for the more concerning hint types, and that outcome-based training to improve it plateaued at low levels. Specific rates are deliberately not reproduced here: they are per-model measurements from one study and would read as current facts about today's models, which they are not. Billing of thinking tokens as output, thinking counting toward the response ceiling, the shift from fixed thinking budgets to per-request decisions, thinking between tool calls, and cache invalidation on configuration change all come from Anthropic's extended thinking documentation, fetched 11 August 2026.

Related → Agents · Prompt caching · Hallucination · All → Concepts