LLM inference and reasoning concept
Reasoning models explained
ReferenceReasoning models are language models that can spend extra work on a prompt before or while producing an answer. They may break a problem into steps, consider different approaches, or work between tool calls. That extra work may help with complex tasks, but a displayed explanation is not necessarily a faithful record of how the model reached its answer. The useful distinction is how the model works during inference, the process of producing an answer from a prompt.
Reviewed 2026-10-09
What makes a model a reasoning model?
A reasoning model is a language model that uses additional computation to work through a request. Some systems represent this work with reasoning tokens; others show a summary or a sequence of steps. These labels describe how a model handles inference, the process of producing an answer from a prompt.
So “reasoning models vs LLMs” is not quite an either-or comparison: reasoning models are a kind of LLM. The practical comparison is with a model or mode that answers with less intermediate work. How much of that work is visible differs across systems, so a displayed explanation should not be assumed to show every internal step.
How the extra work happens
First, the model processes your prompt and works out what the task calls for. It may break the request into parts or consider more than one approach. In some systems, this intermediate work is represented as reasoning tokens; those tokens use space in the model’s context even if the user cannot see them.
Next, the model develops an answer. If tools are available, it may call one, examine the result, and continue reasoning before deciding what to do next. Some systems show thinking as summaries or structured steps; others keep intermediate reasoning hidden. Finally, the model returns an answer, which you can check against the task and any evidence it provides. Extra work can use more tokens and take more time, and it does not guarantee a correct result.
What to be careful about
A step-by-step explanation can sound convincing without faithfully showing how the answer was produced. Researchers have tested whether models mention hints that influenced their answers, then compared those explanations with the hints provided. Their results show why a displayed rationale should be inspected rather than treated as proof of the model’s process. The study tested specific models and evaluation tasks, so its findings should be understood within that scope.
Check the answer itself, too. Look for a missed condition, an unsupported assumption, or a step that does not follow. For arithmetic, logic, or another task with a checkable result, verify it independently. For an open-ended answer, ask what evidence would change your mind and check that evidence directly.
Check your understanding
Try explaining the difference in your own words: a reasoning model is still a language model, while “reasoning” describes additional work it may use during inference. If your explanation treats the two labels as separate kinds of technology, revise it.
Then try a small problem with a checkable answer. Ask a model to solve it and explain its approach. Work out the answer yourself, compare the steps, and note any mismatch. Ask yourself whether the explanation helped you check the answer or whether you accepted it because it sounded clear.
Questions people ask
What are reasoning models?
They are language models that can use additional computation to work through a prompt. Their intermediate work may be hidden, summarized, or shown as steps.
How do AI reasoning models work?
They process a prompt, may work through intermediate steps or approaches, and then produce an answer. Some systems can continue working between tool calls.
Are reasoning models always more accurate?
No. More intermediate work does not make an answer automatically correct. Check the answer and the evidence or steps that support it.
Do reasoning models show their full reasoning?
Not necessarily. A displayed explanation may not faithfully reflect how the model reached its answer.
Learn more
- Reasoning models
Explains reasoning tokens and how they can use context-window space.
- Gemini thinking
Shows how one API represents reasoning with thought steps and summaries.
- Reasoning Models Don’t Always Say What They Think
Explains research testing whether model explanations mention hints that influenced answers.
What changed
Sources
Part of LLM inference and reasoning: Spend inference compute deliberately