Reasoning
Models that think before answering — a large real capability gain, and the visible thinking is not a reliable account of what happened.
Reviewed July 12, 2026Stable
When not to use it
- On easy questions. You pay for tokens that don't help, and the model can talk itself out of a correct answer.
- On retrieval, summarisation, extraction or formatting. There's nothing to reason about.
- As an explanation of the model's behaviour. Turpin et al.: the trace can be a rationalisation that never mentions what actually drove the answer.
- Where answers aren't verifiable. The training method needed a verifier; the capability follows the verifier.
Reach for something else instead
- A cheap model plus routing — decide which questions are hard. That's where the savings are.
- Chain-of-thought prompting — free, works on any model.
- Tool use — a calculator beats reasoning about arithmetic.
- Best-of-n with a verifier — if you can check answers, checking several is often better than thinking harder about one.
Read more on the blog
- How reasoning models work: AI that thinks before answeringIn 2025 a new kind of model arrived: one that pauses to think, working through a problem on internal scratch paper before answering. Reasoning models like o1, o3, and DeepSeek-R1 trade speed for accuracy on hard problems. Here is what they actually do, how they learned to do it, and when the extra thinking is worth it.
- Prompt engineering in 2026: what still works and what changedIf your prompts still open with "act as an expert" and "let's think step by step," you're using 2023 advice on 2026 models, and some of it now makes your output worse. What actually moves results today, what quietly stopped working, and why.
- How AI models are trained: from raw text to a system that helpsA language model isn't programmed, it's grown, in stages, from a firehose of text into a system that answers helpfully. The full modern pipeline: pretraining, supervised fine-tuning, preference alignment, and the reasoning training that defines 2026, what each stage does, why none can be skipped, and how the recipe changed.
- AI scaling laws: why bigger wins, and whether it's endingOne empirical discovery explains most of the last six years of AI and the hundreds of billions spent on it: model performance improves in a smooth, predictable way as you add size, data, and compute. Here is what scaling laws are, why they reshaped the field, the Chinchilla correction, and the live 2026 question of whether pure scaling is running out.
Further reading
- Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models — where the capability gets named.
- Turpin et al. (2023), Language Models Don't Always Say What They Think — chains of thought as post-hoc rationalisation. The essential result.
- DeepSeek-AI (2025), DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — pure RL on verifiable rewards; reasoning behaviours emerging unprompted.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading the trace as an explanation. It's more generated text, from the same process, with the same failure modes.
- Using a reasoning model for everything. It's expensive, slow, and sometimes worse.
- Asking a reasoning model to explain its reasoning. You get a rationalisation of a rationalisation.
- Expecting the maths gains to transfer to judgement tasks. The method needs a verifier and those don't have one.
At a glance
FieldLanguage & LLMs
Mechanismgenerated tokens are computation
How it's trainedRL on verifiable rewards; no human reasoning traces needed
Why maths and code specificallythose are the domains with cheap verifiers
The essential caveatthe trace is not the computation (Turpin et al.)
DifficultyIntermediate
Often compared with
Reasoning model vs. standard model — one spends thousands of tokens computing before it answers and produces a persuasive trace of what it didn't necessarily do.
Where this sits
27 concepts come first. Understanding it opens up 4 more.
14Levelsteps in
27Needs firstconcepts
4Opens up1% of 310
3Areascrosses fields
Learn these firstChain-of-Thought
Computed from the prerequisite graph, not assigned. How this works