The short answer
A reasoning model is a large language model trained to work through a problem in intermediate steps before it gives its final answer. OpenAI's API documentation describes these steps as reasoning tokens: the model uses them to plan, consider alternatives and recover from mistakes, and they are billed as output tokens even though you never see them. In November 2026 most leading models reason, including GPT-6 Astra, Claude Opus 5.5 and DeepSeek-V4-Pro, and you usually control how much with an effort setting.
| Also called | Thinking model, chain-of-thought model |
|---|---|
| How it works | Writes hidden reasoning tokens before the visible answer |
| What you control | Effort level (for example low, medium, high, max) |
| Best for | Multi-step maths, coding, analysis, planning, agent tasks |
| Cost | Slower, and the thinking counts towards usage and billing |
| Examples | GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4-Pro, GLM-5.3 |
What a reasoning model is, in plain English
A normal language model writes its answer one word at a time, straight away. A reasoning model first writes a long internal scratchpad: it breaks the question into parts, tries an approach, checks it and sometimes starts again. Only then does it write the answer you see. TechCrunch's AI glossary, updated in September 2026, puts it simply: reasoning models are developed from traditional large language models and optimised, through reinforcement learning, for this step-by-step style of thinking, which takes longer but tends to give more correct answers, especially in logic and coding.
OpenAI's developer documentation calls the scratchpad reasoning tokens. It says the model uses them to plan, use tools, inspect alternatives and recover from ambiguity, and that they work best for complex problem solving, coding, scientific reasoning and multi-step agent work. Two details matter for anyone paying the bill: the raw reasoning is not shown to you, and it is still billed as output tokens. OpenAI says a single answer can use anywhere from a few hundred to tens of thousands of reasoning tokens.
Where you meet reasoning models in November 2026
Reasoning has gone from a special mode to the default. What changes between tools is how much control you get.
- ChatGPT. OpenAI's GPT-6 family reasons, and since 7 October 2026 GPT-6 can start answering while it is still thinking. Pro plans add a Pro reasoning option powered by GPT-6 Astra, which OpenAI's API guide recommends as the starting point for most reasoning work. See our ChatGPT guide.
- Claude. On Claude Opus 5.5, Anthropic's documentation says thinking is adaptive and always on: the model decides how much to think, and developers steer it with an effort setting of low, medium, high, xhigh or max, with medium as the default. See our Claude guide.
- Gemini. The Gemini app now lets you choose low, medium or high effort for each available model, and Google warns that higher effort uses more of your limit. See our Gemini guide.
- DeepSeek. DeepSeek's models run in thinking mode by default in its API, with low, high and max effort, and the chat app has a thinking toggle. See our DeepSeek guide.
- GLM. Z.ai's GLM-5.3 always reasons. You can choose low, high or max effort, and switching reasoning off is no longer supported.
Why it matters at work
The effort setting is the most useful dial most people never touch. Turning it up is worth it when the task has several steps that depend on each other and a wrong step ruins the result: reconciling two spreadsheets, checking a contract against a policy, debugging code, planning a project with constraints, or any task you hand to an AI agent. Turning it down is better for rewriting an email, summarising a short document or brainstorming names, where extra thinking adds waiting time and spends your plan's allowance without improving much.
There is also a budgeting point for teams that use an AI API. Because hidden reasoning is billed as output, the same question can cost very different amounts at different effort levels. OpenAI recommends reserving at least 25,000 tokens for reasoning and output when you start experimenting, so answers are not cut off before the visible part begins. Our guide on how to choose the right AI model covers the trade-off by task.
An example you can try today
Take a problem with a trap in it, for example this one, and run it twice in the same assistant: once with the lowest effort or fastest model, once with the highest effort your plan offers.
- Paste the first prompt below into ChatGPT, Claude, Gemini or DeepSeek with low effort or the fast option.
- Start a new chat, switch to high effort or the thinking option, and paste it again.
- Compare the two answers. Check the arithmetic yourself with a calculator.
- Note how much longer the second one took. That waiting time is the cost you pay for reasoning on every task.
The point is not which answer is right this time. It is to get a feel for which of your everyday tasks need the slow setting and which do not.
Common misconceptions about reasoning models
- "More thinking is always better." No. For simple tasks it mostly adds delay and cost. Google itself notes that higher effort in Gemini uses more of your limit.
- "The thinking I see is everything the model thought." Often it is a summary. OpenAI says it does not expose raw reasoning tokens and offers a summary instead, and Anthropic omits thinking content by default on recent Claude models unless developers opt in.
- "Reasoning models do not make things up." They make fewer errors on multi-step problems, but they still hallucinate. Check figures, quotes and sources, as our guide on checking if AI is telling the truth explains.
- "A reasoning model is a different kind of AI." It is still a large language model. It has been trained to spend more of its output on working before answering.
- "I can make it more careful with temperature." On many current reasoning models the temperature setting is gone: Claude Opus 5.5 does not accept it, Gemini 4 ignores it and OpenAI's GPT-6 models reject it whenever they reason. Effort is the dial to use.
Related terms
- Large language model (LLM): what a reasoning model is built from.
- Token: the unit reasoning is measured and billed in.
- Context window: reasoning tokens take up space in it.
- Inference: running the model, which reasoning makes longer.
- Deep research: a reasoning model plus many web searches.
- Benchmark: the tests labs use to claim reasoning gains.
For a side-by-side view of how the big assistants handle hard problems, see ChatGPT vs Claude and ChatGPT vs Gemini.
Prompts to copy
Test low vs high effort on a trap question
A team of 7 people each work 37.5 hours a week. Two of them are part-time at 60%. Holidays remove 3 weeks a year for everyone. The project needs 9,000 person-hours. Starting on Monday 5 January 2026, in which week does the project finish if everyone works only on it? Show your working step by step and state any assumption you make.
You get: A worked answer you can compare across effort levels
Ask the model to plan before it answers
Before you answer, list the steps you will take to solve this and the main ways the answer could go wrong. Then solve it, check your result against each risk you listed, and give the final answer with a one-line confidence note. Problem: [describe your multi-step problem, for example reconcile these two supplier price lists and flag every mismatch].
You get: A more careful answer with its checks visible