tokenkeret

in English: token budget / thinking budget

advancedwidespread

The token limit is a predefined upper bound on how much text an AI model or agent can process and how much it can "think" on a task.

In plain terms

AI models measure work in tokens: one token is roughly a word fragment. Every token read, processed, and written costs time and money. The token budget determines how much the system can spend on a task.

It is like telling a mechanic: you can work on the car for at most two hours, then tell me what you found. You do not specify what to do, but how much effort can be spent.

The thinking budget is a narrower form of this. It limits how much the model can reason internally before responding. A simple question requires little, while complex analysis or programming needs more.

For agents, the framework is especially important because the agent works independently in many steps: it searches, reads files, and calls tools. Without a framework, a stuck agent could spin the bill for hours.

How it works

The developer or the service sets an upper limit: a number (for example, 10 000 tokens for thinking) or a level (low, medium, high).
The model consumes the limit while working: tokens count for what it reads, what it thinks through internally, and what it finally writes.
In an agent, every step proceeds from the context: tool results, histories of previous steps, and the work of any helper agents.
If the budget runs low, a well-built system shortens, summarizes, or closes the task and indicates how far it has progressed. A poorly built one simply halts abruptly.
Some newer models decide themselves how much thinking is needed for a given question, within the given limit.

An everyday example

You ask a chatbot to compare three mortgage offers and enable the "extended thinking" mode. The response comes slower but is more thorough because the model has been given a larger thinking budget. However, if you see that you have reached the limit in the middle of a long task, it is usually not the task's token budget but your subscription's periodic usage cap: a related concept, but not the same.

Who is affected

developerIt sets the framework, thereby balancing response quality, speed, and cost. → Code

entrepreneur, executiveThe token limit sets an upper bound on how much it costs to operate an agent and protects the company from runaway bills.

employeeIf you work with an AI tool, the budget allocated to the task determines how deep a piece of work you can assign to it. However, how much you can use it in total over a day or a week is decided by your subscription's usage limit. → Work & HR

everyoneThe 'fast' and 'thinking' modes behind chatbots are usually this setting, displayed as a button. Sometimes the button switches to another model.

A common misconception

MythThe larger the token limit, the better and more reliable the answer.

In realityMore thinking helps with difficult, multi-step tasks; for simple questions, it only slows things down and increases costs. The model can make mistakes even with a large thinking budget.

Limits and risks

The cap is an upper limit, not a promise: the model does not necessarily use it fully and does not guarantee quality. In some systems, it is only a guideline that the model may slightly exceed.
Due to overly narrow constraints, the response or the agent's work may be interrupted, and this is not always clearly detectable.
You pay for tokens consumed in reasoning even if you do not see the line of thought or only see a summary.
Providers calculate differently and name the setting differently, so the limits cannot be compared directly.

Similar, but not the same

context window – The context window shows how much text the model sees at once, while the token limit shows how much it can use for the task.
rate limit – Usage limits are set by the provider per time unit (minute, day, week) to prevent overload or as part of a subscription. The token limit, however, is set by you or the developer for a specific task.
maximum response length – This limits the maximum amount of text the model can generate in a single response; at several providers, thinking tokens also count. The token budget is broader: it can regulate thinking separately, and in the case of an agent, the entire multi-step session.

Related concepts

Broader

resource management, cost control, agent control

Narrower

thinking framework, feladatkeret, reasoning effort

On the edge

token, context window, reasoning model, usage limit, kontextusmenedzsment

A brief history

Token-based limits appeared alongside the programming interfaces of large language models. OpenAI already offered adjustable reasoning effort levels for its reasoning models by the end of 2024. Anthropic introduced a thinking budget specified in tokens in February 2025 with the Claude 3.7 Sonnet model, and Google added a similar setting to its Gemini 2.5 models later that same year.

Where things stand

At the end of 2026, every major model provider offers some form of reasoning or task-level framework. Instead of specifying an exact token count, tiers and approaches in which the model itself judges how much reasoning is needed are becoming more common. For some newer models, an exact token count can no longer be set at all.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Copilot: credit complaints, token budget and new Lite and Balanced tiersAug 4, 2026 · Code

Also known as

tokenkeret · token-keret · token budget · thinking framework · token budget · thinking budget · reasoning budget · budget_tokens

Sources

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