reasoning effort

advancedwidespread

A setting that determines how much the AI 'thinks' before answering: more thinking yields a more thorough but slower and more expensive response.

In plain terms

Reasoning effort is like telling a specialist how much time to spend on your question. A quick phone answer can be good, but for a complex case you ask them to sit down, work it out, and review it once more.

Some current language models write an internal 'draft' before responding: they break down the task, weigh options, and self-check. The reasoning effort regulates how large this draft can be. Usually, you can choose from a few levels, such as low, medium, or high. In some systems, you specify this with a number indicating the maximum amount of thinking allowed.

The setting does not give the model more knowledge: the same model works, it just puts more or less effort into the answer. With simple questions, a high setting is wasteful, while with difficult tasks, a low setting can produce superficial or erroneous results.

How it works

You or the application developer select a level (for example, low, medium, high) before the query is sent to the model.
The model adjusts how much it thinks based on the level: at a higher level, it breaks the task into parts for longer, weighs more options, and checks more. The model learns this behavior during training, so the effect is only a rough guide, not precisely predictable.
This thinking is also text generated token by token by the model. You usually do not see it, or only in summary, but it costs time and money.
At the end, the model writes the visible answer, which for more difficult tasks is typically more accurate and well-founded after longer reasoning.
For agents, i.e. AI programs that work independently in multiple steps, the setting also influences how thoroughly they plan and how many tools they call during a task.

An everyday example

In a chat program, alongside your question, you can choose between a 'fast' and a 'thinking' or 'thorough' mode. If you are only curious about the meaning of a word, the fast mode answers in moments. If you want to compare the terms of a credit offer, the thorough mode works longer but typically makes fewer mistakes.

Who is affected

everyoneIf you use an AI chatbot, the choice between "fast" and "thinking" modes largely represents this setting. In some services, a different model may take over the work in the background.

developerChoosing the level balances response quality, waiting time, and cost. → Code

entrepreneur, executiveThe bill for corporate AI use is directly affected by whether routine tasks are running at unnecessarily high levels.

parent, student, teacherFor mathematical or logical tasks, a higher reasoning level typically gives a more reliable solution than a quick answer. → Education & Family

A common misconception

MythIt is always best to use the highest setting because it provides the best answer.

In realityFor simple tasks, more thinking does not improve the answer; it only becomes slower and more expensive. The level should be adjusted to the task's difficulty.

Limits and risks

High reasoning levels are no guarantee: the model may still err or hallucinate facts after prolonged thinking, so verify important claims.
Thinking itself consists of paid tokens, so a higher tier can multiply both the cost and the waiting time.
The names and effects of tiers vary by provider and model, so the "high" level of one system cannot be directly compared to that of another.
The displayed reasoning is often just a summary and does not necessarily reflect exactly how the model arrived at the answer.

Similar, but not the same

temperature – Temperature regulates how varied or predictable the text is, not how much the model thinks beforehand.
chain of thought – Chain of thought is the method of step-by-step reasoning itself, while reasoning effort is the control that determines how much of it is used.
model size – A larger model 'knows' more, while reasoning effort adjusts how much work is spent within the same model.

Related concepts

Broader

large language model, inference, model parameters

Narrower

thinking framework, extended thinking, adaptive thinking

On the edge

reasoning model, chain of thought, token, temperature, AI-agent

A brief history

Reasoning models that 'think' step by step reached the general public in autumn 2024 (OpenAI o1). In December 2024, OpenAI made a setting called 'reasoning effort' available to developers. In 2025, Google and Anthropic also introduced similar controls, initially as a 'thinking budget', later with preset levels as well.

Where things stand

At the end of 2026, most leading language models let you adjust how much they think, and in chat applications this appears as a simple mode selector. It is increasingly common for the model itself to assess how much thinking a task requires.

Status: 2026 November

In the news

Appears in 13 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Norwegian media companies sue Meta over use of their journalism for AIOct 27, 2026 · Jog
Siemens and Nvidia: engineering effort cut by 85% at Erlangen plantOct 19, 2026 · Ipar
Google says for AI content in SEO, effort matters, not the toolOct 16, 2026 · Marketing
Banking watchdogs step up AI scrutiny amid financial stability fears, sources sayOct 12, 2026 · Jog
HHS and ARPA-H move to speed up clinical trials with AIOct 5, 2026 · Health
Copilot code review gets API support and a new default effort levelOct 2, 2026 · Code
Colorado mother warns of AI chatbot risks to teenagersSep 30, 2026 · Education & Family
Intel to push AI and STEM education in Israeli schools, Education Ministry saysSep 25, 2026 · Education & Family

Also known as

reasoning effort · reasoning_effort · effort · thinking budget · thinking level · reasoning effort · thinking level · reasoning level

Sources

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