test-time computation (TTC)

in English: test-time compute

advancedwidespread

This is the computational work the language model performs while answering your question: the more it can 'think', the better the answer may be.

In plain terms

A language model has two major phases in its life. In the first, it is trained: this is like school years. In the second, it is used: this is the exam, when you ask and it answers. Inference-time computation is the work the model performs in this second phase while processing your question.

Think of a chess player. Their knowledge is the same in a blitz game and a long match, yet they play better if they have ten minutes per move, because they can calculate more variations. Similarly, a model can be more accurate if it does not immediately blurt out the answer, but first thinks through step by step, tries multiple solutions, and checks itself.

This is important because previously, smarter AI could mainly be built with a larger model and more training data. Test-time computation offers another path: the same model can deliver better results with more thinking time. The cost is that the response is slower and more expensive.

How it works

You ask a question, and the system decides, or you set how much "thinking budget" the model gets for the answer.
The model does not immediately write the final answer; instead, it first prepares an internal working text: it breaks the task into parts and proceeds step by step.
On a hard task, it can try several solution paths, discard the dead ends and check partial results, even by running code or searching.
In the end, it formulates the solution deemed best as its response; from long reasoning, you often see only a brief summary.
Thinking is also computation measured in text units (tokens), so you wait longer for the response and must pay for it: in developer usage, the provider bills per token, while in subscriptions, it consumes your usage quota.

An everyday example

In a chat program, you request a complex loan comparison and turn on the 'thinking' or 'thorough' mode. The answer does not arrive in two seconds but in half a minute, during which you see: 'Thinking…'. In return, the calculation is more likely to be error-free.

Who is affected

everyoneThe choice between fast and thorough modes in chat programs is actually about how much test-time computation you request.

entrepreneur, executiveThe cost of using AI increasingly depends on how much the model thinks about each task, not just on which model you choose.

developerYou must configure how much reasoning budget your application or agent receives for each task, as this determines quality, speed, and the bill. → Code

parent, student, teacherFor mathematical and logical tasks, the reasoning mode is noticeably more reliable, but the derivation should still be verified. → Education & Family

A common misconception

MythIf the model thinks longer, its answer will definitely be correct.

In realityMore thinking improves the odds but does not replace missing knowledge: the model can still be wrong after a long, convincing derivation. For simple questions, it typically adds little, but slows things down and increases costs.

Limits and risks

More reasoning means more waiting, higher costs, and greater energy consumption.
The benefit diminishes beyond a certain point: ten times the computation does not yield ten times the better answer.
What the model does not know or knows poorly, it cannot invent even with prolonged deliberation; fresh or missing information requires a source, not time.
The displayed 'chain of thought' is often just a summary and does not necessarily reflect exactly how the model arrived at the answer.

Similar, but not the same

training – Training is a one-time, prior learning process that shapes the model's knowledge, while test-time computation is work performed anew for each question.
chain-of-thought – Chain of thought is a specific method for step-by-step reasoning, while test-time compute is the broader principle that includes search, multiple attempts, and self-checking.
fine-tuning – Fine-tuning changes the model itself, while test-time computation leaves the model untouched, only allowing it more work before generating a response.

Related concepts

Broader

inference, large language model, computing capacity

Narrower

chain of thought, extended thinking, thinking framework, self-check

On the edge

reasoning model, token, training, AI-agent, scaling

A brief history

The idea originates from game programs: chess and Go programs have long searched for the best move during play. DeepMind's AlphaGo program defeated Lee Sedol in 2016 by combining trained neural networks with search during the game. For language models, the chain-of-thought method described in 2022 laid the groundwork, and it became widely known through OpenAI's o1 model, presented in September 2024 (initially as o1-preview).

Where things stand

At the end of 2026, almost all leading language models offer an adjustable reasoning level, and choosing between a fast and a thorough mode is now familiar in chatbots too. The performance of agents that work autonomously for long stretches rests largely on this, so the centre of gravity of compute demand is shifting from training towards usage.

Status: 2026 November

Also known as

test-time compute · inference-time compute · test-time scaling · inference-time scaling · inference-time computation · runtime computation · thinking time · 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.