thinking token

in English: reasoning tokens / thinking tokens

advancedwidespread

A thinking token is the text unit a language model uses before the final answer to reason through the task step by step, for which payment is usually required.

In plain terms

When you write a math test, the final answer is accompanied by a draft sheet: you calculate, try things out, and cross them out there. Thinking tokens are the draft sheet of a language model. The model first works through the reasoning 'in its head' and only then formulates the answer you see.

A token is a small unit of text (roughly a word fragment); the model's work and its price are measured in these. Thinking tokens are the same kind of units, but they go into the pondering preceding the answer rather than into the answer itself. This makes the model better at multi-step tasks: calculation, planning, programming, and complex decisions.

You pay for a more accurate answer with slower operation and higher costs. For a simple question, prolonged thinking is wasteful, but for a difficult task, it pays off. Therefore, in many services, you can set how much the model should think, while in others, the model decides based on the difficulty of the task.

How it works

The model receives your question and decides (or is instructed by settings) how much prior thinking is needed.
Before answering, it writes internal text: it breaks down the task, tries out solution paths, and checks and corrects its own steps. This text consists of thinking tokens.
Relying on this internal text, it writes the final answer, which now consists of proper output tokens.
The provider either hides the reasoning process or shows only a brief summary, while in some systems it can be read in full.
Billing typically counts reasoning tokens as output tokens, and they also take up space in the amount of text the model can handle at once (the context window).

An everyday example

You ask the chatbot which loan offer is better, and the screen displays for half a minute: "Thinking...". At this moment, the model uses thinking tokens: it calculates all variations in the background before giving an answer. If a company calls this from a program, this invisible part also appears on the bill.

Who is affected

developerYou must set how much the model can think, as this simultaneously affects quality, response time, and cost. → Code

entrepreneur, executiveThe AI service bill can spike even if responses are short, because you pay for invisible thinking as well.

employeeIf you know when to switch on the "thinking" mode, you get better answers for difficult tasks and faster ones for easy ones. → Work & HR

everyoneThe process behind the "Thinking..." message is why you sometimes have to wait longer for a response.

A common misconception

MythThe displayed reasoning shows exactly how the model arrived at the answer.

In realityThe written chain of thought is a useful pointer but not a faithful transcript: research suggests that factors not included in the described steps can influence the model's response. Often, you only see a summary prepared afterwards.

Limits and risks

More thinking is no guarantee of a correct answer: a model can still be wrong after long deliberation, just more convincingly.
Thinking tokens increase cost and waiting time, making them unnecessary for simple questions.
You often cannot see the reasoning process, or only see a summary, making it difficult to verify where the paid resources were spent.
If reasoning consumes too much of the context window, the final answer may be truncated or never generated.

Similar, but not the same

kimeneti token – Technically, the thinking token is also output generated by the model and counts toward the bill. The difference is that it does not appear in the response shown to you but belongs to the internal work preceding the response.
chain-of-thought – Chain of thought is the method itself, meaning step-by-step reasoning, while a thought token is the unit of measurement used to gauge its volume and cost.
context window – The context window is the model's entire 'desktop', and reasoning tokens are just one thing occupying space on it.

Related concepts

Broader

token, large language model, inference

Narrower

thinking framework, reasoning summary

On the edge

chain of thought, reasoning model, kimeneti token, context window, token-based pricing

A brief history

Step-by-step reasoning, the idea behind chain-of-thought, was described by Google researchers in 2022. Separately billed 'reasoning tokens' first appeared in September 2024 with the OpenAI o1 model. From the end of 2024 and throughout 2025, other major developers introduced their own versions: Google with its Gemini 'thinking' models, DeepSeek with R1, and Anthropic with 'extended thinking.'

Where things stand

By the end of 2026, almost all leading language models can think before answering, and the extent is mostly adjustable, or the model itself adjusts it to the task's difficulty. In coding and other independently working agents, these tokens can account for a significant part of the cost.

Status: 2026 November

Also known as

thinking tokens · reasoning token · inference token · reasoning token · reasoning tokens · thinking token · thinking tokens · extended thinking

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.