Products, laws and numbers here change fast. Checked October 2026.
In plain words
Ask someone what 17 times 24 is and they will not shout a number. They work it out first. Early chatbots did shout: they started writing an answer immediately, word by word, with no chance to think it through.
A reasoning model is trained to think first. Before replying, it writes out a long chain of working: trying an approach, checking it, noticing mistakes, trying again. Then it gives a final answer. Most apps hide this working or show only a short summary.
The extra thinking costs time and money, because every step is more text the model has to produce. In return, reasoning models do far better on maths, coding, planning and tricky logic. For a simple question, the thinking adds little.
As of October 2026, the main models from OpenAI, Google and Anthropic all reason, and adjust how long they think to the difficulty of the question.
Everyday example
Rough work in an exam
In a maths exam, the margin of the answer sheet fills with rough work: a first try, a crossed-out line, a check. The examiner mostly looks at the final boxed answer. A reasoning model does the same: it fills a hidden margin with working, then writes the clean answer. More rough work on harder questions usually means fewer careless mistakes.
Where the example breaks: A student's rough work shows how they really solved it. A model's written thinking is not always a faithful record: research in 2025 found models often used hints without mentioning them in their reasoning.
How it actually works
Reasoning models are still large language models. What changed is how they are trained and how much they write before answering.
Thinking is just more tokens
The model's "thoughts" are tokens it generates and then reads back, the same next-token prediction as always. Writing intermediate steps gives it more computation per answer. This idea started as a prompting trick called chain of thought.
Trained with reinforcement learning
Models are set problems with checkable answers, such as maths or code with tests, and rewarded when the final answer is right. Through reinforcement learning they learn which thinking habits pay off, like double-checking or backtracking. DeepSeek showed in January 2025 that this can work without human-written examples of reasoning.
Test-time compute
OpenAI reported that its first reasoning model got better both with more training and with more time spent thinking on each question. Spending more compute at the moment of answering became a new way to improve results.
Effort settings
Developers can usually choose an effort level. OpenAI's API, as of October 2026, offers levels from none or minimal up to max; lower effort favours speed and fewer tokens. Anthropic's newer Claude models decide for themselves whether and how much to think, guided by an effort setting.
Hidden thinking still counts
Reasoning tokens use up the context window and are billed like output, even when they are not shown.
- First widely released
- OpenAI o1-preview, 12 September 2024
- Open weights
- DeepSeek-R1, January 2025, with its thinking shown in full
- One model, two modes
- Anthropic's Claude 3.7 Sonnet, February 2025: "both an ordinary LLM and a reasoning model in one"
- Exam jump
- On the 2024 AIME maths exam, GPT-4o solved about 12% and o1 about 74% on a single try (OpenAI)
Questions people ask
What is the difference between a reasoning model and a regular LLM?
A regular LLM starts answering straight away. A reasoning model first generates a long hidden working-out, then answers. Underneath they are the same kind of model; the reasoning one has been trained to use that thinking step well, and costs more per answer.
Do reasoning models actually think?
They produce step-by-step text that works like thinking and clearly improves results. Whether it is thinking in the human sense is debated. Research also shows the written reasoning does not always reflect what drove the answer, so it should not be read as a perfect window into the model.
When should we use a reasoning model?
For multi-step problems: maths, code, analysing data, planning, comparing options with many conditions. For quick facts, rewriting or casual chat, a fast mode is cheaper and just as good. Newer models often judge this for themselves.
Why can't we see the full reasoning?
Some companies hide the raw thinking and show a summary. OpenAI cited safety monitoring, user experience and competition when it launched o1. Others, such as DeepSeek-R1, show it all.
Are reasoning models less likely to hallucinate?
They make fewer mistakes on problems they can check step by step, like arithmetic. But they can still state false facts confidently, and long reasoning can wander. Checking important answers is still wise (hallucination).