What is test-time compute?

Spending more computation when the question is asked rather than only during training. Longer reasoning, several attempts scored against each other, or a search over candidate answers. It is a second way to buy quality that does not require training a bigger model.

Can I see what the model is thinking?

Usually only a summary. Labs withhold the raw trace for competitive and safety reasons, and the shown text is a cleaned version. Useful for following the shape of an approach, not reliable as proof of how the answer was reached.

When should I not use a reasoning model?

When the task is retrieval, extraction, formatting or short chat. You pay for the extra thinking tokens and wait longer for an answer that a fast model gets right anyway. Save it for problems where a wrong intermediate step ruins the result.

Are reasoning models trained differently?

Yes. After ordinary pretraining they get a phase of reinforcement learning on problems with checkable answers, such as math and code, where the model is rewarded when its final answer is right. That teaches it to spend tokens on working-out that actually improves the answer, rather than just writing more.

Does a reasoning model cost more to use?

Yes, in tokens and time. The hidden reasoning tokens are generated and billed like any output, so a hard question can cost several times a plain answer and take much longer. Products expose a thinking budget or a toggle so you pay for reasoning only where it helps.

Plain-language explainer

Reasoning models, or why some answers take longer

A reasoning model is a language model trained to write out a long, mostly hidden chain of intermediate steps before it commits to an answer. It generates those steps token by token like any other text, which lets it decompose a problem, try an approach and correct itself. It trades more time and tokens for better results on hard problems.

How do reasoning models actually work?

A reasoning model writes out a long private working-out before it answers. That draft is generated the same way as any other text, one token at a time, but it gives the model somewhere to break a problem into steps, try an approach and correct itself. Spending more tokens at answer time reliably improves hard maths, code and multi-step logic. It costs more and takes longer, and on simple questions it buys nothing.

Last reviewed September 6, 2026

Reading is the slow way. Start with a free lesson you can operate right now.

Start free: From the internet to your answer →

Free, no code, no signup.

Then go deeper: The model thinks before it answers Locked

What people get wrong

The model is thinking in some new way. It is generating tokens exactly as before, just generating far more of them before it commits to an answer.
The visible summary is the real reasoning. Most products show a cleaned-up sketch, not the raw trace, so it is a rough guide and not an audit trail.
Reasoning mode is always better. On lookup and formatting tasks it adds latency and cost for no gain, which is why products let you switch it off.

Where you see it in real products

The thinking or extended reasoning toggle in a chat assistant.
Per-answer effort settings in developer APIs, priced by the tokens the model thinks with.
Coding tools that plan a multi-file change before touching any file.

Common questions

What is test-time compute?: Spending more computation when the question is asked rather than only during training. Longer reasoning, several attempts scored against each other, or a search over candidate answers. It is a second way to buy quality that does not require training a bigger model.
Can I see what the model is thinking?: Usually only a summary. Labs withhold the raw trace for competitive and safety reasons, and the shown text is a cleaned version. Useful for following the shape of an approach, not reliable as proof of how the answer was reached.
When should I not use a reasoning model?: When the task is retrieval, extraction, formatting or short chat. You pay for the extra thinking tokens and wait longer for an answer that a fast model gets right anyway. Save it for problems where a wrong intermediate step ruins the result.
Are reasoning models trained differently?: Yes. After ordinary pretraining they get a phase of reinforcement learning on problems with checkable answers, such as math and code, where the model is rewarded when its final answer is right. That teaches it to spend tokens on working-out that actually improves the answer, rather than just writing more.
Does a reasoning model cost more to use?: Yes, in tokens and time. The hidden reasoning tokens are generated and billed like any output, so a hard question can cost several times a plain answer and take much longer. Products expose a thinking budget or a toggle so you pay for reasoning only where it helps.

Related explainers

Next-token prediction Inference economics LLM limits

One idea at a time, in your inbox

New lessons and explainers, written the way these pages are. Now and then, not daily, and never a sales sequence.

We email you a confirmation link first. Unsubscribe in one click, any time. Privacy.

Part of See How AI Works, a free interactive course, where you learn how modern AI works by operating it, not watching videos.