Insight · October 12, 2026
Think first,
answer second.
A reasoning model spends compute to think in steps before it answers. What it is, how it differs from a standard model, and when the extra time and tokens are worth it.
01 · In one paragraph
What a reasoning model actually is.
A reasoning model is a large language model trained to work through a problem in steps before it answers, rather than reply in one pass. It spends extra computation at the moment you ask, called test-time compute, to write out a chain of thought and check its own work. That lifts its accuracy on hard math, code, and analysis.
The model you get from a lab in its plain form answers from pattern, fast and in one go. A reasoning model takes the same kind of network and trains it to slow down on hard questions, laying out intermediate steps the way a person works a proof on paper. The steps are the point. They are where the model catches its own mistakes before they reach you.
02 · One pass versus many steps
A standard model replies. A reasoning model deliberates.
Ask a standard model a question and it produces an answer in a single forward pass. That is fast and cheap, and for most questions it is enough. The weakness shows on problems where one wrong early step quietly ruins the rest: a long calculation, a tricky edge case in code, a chain of logic with several links.
A reasoning model is built for exactly those. Anthropic describes its extended thinking as letting the very same model give itself more time and expend more effort before it answers, rather than switching to a separate model. The thinking is serial test-time compute: sequential reasoning steps, each adding a little more computation, until the model settles on a reply.
This is a different lever from the ones most teams already know. It is not a bigger prompt and not more training data. For how a model reads what you give it, see context engineering. A reasoning model changes what the model does after it has read, during the answer itself.
The one line to keep
“A standard model answers from reflex. A reasoning model stops to think, and you pay for the thinking in time and tokens.”
03 · Where the thinking comes from
The habit is trained, mostly by reward.
A reasoning model is not just a standard model with a longer prompt. The habit of thinking in steps is trained in, and the method that made it work at scale is reinforcement learning. The model tries a problem, is rewarded when it reaches the right answer, and over many rounds learns the reasoning that tends to get there.
DeepSeek made the clearest public demonstration. In January 2025 it released DeepSeek R1, a family of open reasoning models, and showed that its R1 Zero variant learned to reason through large scale reinforcement learning with no supervised fine-tuning first. DeepSeek calls it the first open research to validate that reasoning can be incentivized purely through reinforcement learning. Left to practice against a reward, the model developed self checking, reflection, and long chains of thought on its own.
The chain is literal. DeepSeek recommends that the model begin every answer with a think tag, so the reasoning runs first and the reply follows. DeepSeek R1 reached performance comparable to OpenAI o1 across math, code, and reasoning, and shipped under the MIT license alongside six smaller models distilled from it, from 1.5 billion up to 70 billion parameters.
04 · One measured example
More thinking, measurably better answers.
The payoff is not a feeling, it is a curve. Anthropic reported that on the 2024 American Invitational Mathematics Examination, Claude 3.7 Sonnet's accuracy rose with the number of thinking tokens it was allowed to spend, improving in a predictable way as the budget grew. More thinking, more correct answers, up to a point.
A second method pushes further. Running the model many times in parallel and keeping the best result, with the equivalent of 256 independent samples, a learned scoring model, and a 64K token thinking budget, Claude 3.7 Sonnet scored 84.8% on GPQA, a hard set of graduate level biology, chemistry, and physics questions, including a physics subscore of 96.5%. The answers get better because the model can explore several lines of thought and keep the one that holds.
Serial test-time compute · one long chain · accuracy rises with the budget Parallel test-time compute · 256 samples, 64K · GPQA 84.8%, physics 96.5% More compute at answer time, not more training.
Source · Anthropic, Claude 3.7 Sonnet extended thinking
05 · The dial you control
You decide how hard it thinks.
Thinking budget
Set a token ceiling
Claude 3.7 Sonnet lets you tell the model to think for no more than N tokens, for any value up to its output limit of 128K tokens. More budget for the hard questions, less for the easy ones.
Effort level
Trade capability for speed
Claude Opus 4.5, released November 2025, adds an effort parameter so you can minimize time and spend or maximize capability. At a medium effort it matched its own best score on SWE-bench Verified while using 76% fewer output tokens.
Hybrid mode
One model, two gears
Anthropic calls Claude 3.7 Sonnet both an ordinary model and a reasoning model in one, so you pick when it answers at once and when it thinks longer first. The reasoning is a mode, not a separate product.
The dial matters because thinking is not free. A short question answered at full effort wastes tokens and time. The skill is matching the depth of thought to the difficulty of the task, and the controls above are how you do it per call.
06 · When it earns its keep
Reach for it where a mistake is expensive.
A reasoning model earns its extra cost on work where being wrong is worse than being slow. Hard math and data analysis, code that has to pass real tests, planning that spans many steps, and decisions where you want to see the working, not just the verdict. These are the tasks where a few extra seconds and a larger token bill buy a materially better answer.
The cost is real and worth naming. Every thinking token is a token you pay for and wait on. That is why the newest models pull in the other direction too. Opus 4.5 reaching its best result with 76% fewer tokens at a medium effort is the frontier learning to think only as much as a problem needs. For teams building on these models, that efficiency is the difference between a feature that ships and one that is too slow or too expensive to run.
Reasoning also sits underneath the systems getting the most attention now. An AI agent that plans, calls tools, and recovers from errors is leaning on a reasoning model to decide each next step. The better the model reasons, the less the agent flails.
07 · When not to reach for it
It is not a better model for everything.
Do not reach for a reasoning model on simple, high volume work. A classification, a short rewrite, a lookup, a format change: a standard model answers these instantly and for a fraction of the cost. Paying for a chain of thought on a one line task is pure waste, and at scale the waste is the whole bill.
Do not treat the visible thinking as the truth. Anthropic is direct that the shown reasoning may not be faithful: its own research suggests models often make decisions based on factors they do not state in their thinking. The trace is useful for catching errors and for trust, but it is not a guaranteed window into why the model did what it did. Read it as a helpful artifact, not as proof.
And more thinking is not always more right. The gains flatten as the budget grows, and on some tasks they stop. Treat the thinking budget as a tuning decision you measure, the same way you would any other, rather than a setting you turn to the maximum and forget.
Closing
Pay for thinking only when it earns it.
A reasoning model is a standard model taught to slow down and work in steps, and to let you set how long it takes. Point it at the problems where a wrong answer costs more than a slow one, keep it off the easy work, and tune the budget to the task. Named the right way, the dial is the whole advantage.
Written by · October 12, 2026
Donny Smith · ECD, Founder, Bttr.
Over the past 15 years, he has led creative teams and contributed to products used by millions of people worldwide, working with companies including Apple, Opendoor, JP Morgan, GE Aerospace, Pepsi, and Alterra Mountain Company.
LinkedInShare this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.