What Is a Reasoning Model in AI?
Reasoning models think before they answer. Here is how they work, how they differ from regular AI models, and when the extra wait is worth it.
Try a reasoning modelQuick answer: a reasoning model is an AI language model that works through a problem step by step before it writes its final answer. That hidden "thinking" makes it slower and more expensive per answer than a regular model, but noticeably better at math, coding, planning, and anything with several steps that depend on each other. For quick questions and everyday writing, a fast non-reasoning model is usually the better choice.
Quick picks: what are you trying to do?
- Debug code or solve a math problem
- Use a reasoning model with higher effort
- Draft an email or summarize a page
- A fast, non-reasoning model is enough
- Understand the difference
- See the comparison table
- Pick the right effort level
- Low, medium or high reasoning effort
Quick Takeaways
- Reasoning models plan, try approaches, and check their work before answering; regular models start answering immediately.
- The thinking costs time and tokens, so reasoning is worth it for hard problems and wasteful for easy ones.
- Most major model families now offer a reasoning mode: OpenAI, Claude, Gemini, DeepSeek, Grok, and Kimi all have thinking variants.
- "Reasoning effort" lets you choose how much thinking a model does, trading speed for depth.
What is a reasoning model?
A standard large language model writes its answer one token at a time, straight away. Whatever it "thinks" happens inside that single pass. A reasoning model adds a step in front: before the visible answer, it generates a chain of intermediate reasoning, often called thinking tokens or reasoning tokens. In that space it breaks the question into parts, tries an approach, notices when something does not add up, and backtracks.
A useful comparison is a student showing their work on a math problem. The final answer is shorter than the working, but the working is what makes the answer right. Reasoning models are trained, largely with reinforcement learning on problems with checkable answers, to produce working that leads to correct results more often.
Many apps hide the raw reasoning and show only the final answer, or a short summary of the thinking. Either way, you pay for it in waiting time, and on paid APIs the reasoning tokens are usually billed as output tokens even when you never see them.
How do reasoning models work?
Under the hood, a reasoning model is still a language model. What changes is how it is trained and how much it is allowed to generate before answering:
- Read the problem and plan. The model restates what is being asked and lists the steps or constraints.
- Work through the steps. It calculates, writes draft code, or compares options, one piece at a time.
- Check and revise. It tests the draft against the question, catches contradictions, and tries another route if the first one fails.
- Write the final answer. Only then does it produce the response you read, usually shorter and more direct than the reasoning behind it.
The longer a model is allowed to think, the more of this checking it can do. That is why the same reasoning model can give a better answer at a higher effort setting, and why hard problems take longer.
Reasoning vs non-reasoning AI models
| Non-reasoning (standard) | Reasoning (thinking) | |
|---|---|---|
| How it answers | Starts writing the answer right away | Thinks step by step, then answers |
| Speed | Fast, near-instant | Seconds to minutes, depending on effort |
| Cost per answer | Lower | Higher: thinking tokens add up |
| Best at | Chat, drafting, rewriting, summaries, brainstorming | Math, coding, debugging, planning, multi-step analysis |
| Typical failure | Confident answer that skipped a step | Overthinks simple questions, slow |
| Prompting style | Benefits from "think step by step" and examples | Give the goal and constraints; it plans on its own |
The line is blurrier than it used to be. Many current models are hybrids: the same model can answer instantly or think first, depending on a setting. So the practical question is less "which model is a reasoning model?" and more "should this request use reasoning, and how much?"
Examples of reasoning models
Reasoning went mainstream when OpenAI released o1 in 2024, followed by open models such as DeepSeek R1 in early 2025. Since then, almost every major model family has added a thinking mode:
- OpenAI (GPT): reasoning models with adjustable reasoning effort, now built into the main GPT line.
- Anthropic (Claude): Claude models that can answer directly or use extended thinking first.
- Google (Gemini): Gemini models with built-in thinking, often with a readable thinking summary.
- DeepSeek: thinking variants alongside fast chat models.
- Others: Grok, Kimi K2 Thinking (Moonshot AI) and GLM (Z.AI) also ship reasoning models.
On Chocolatey AI you can switch between models from these providers in one app, including GPT, Claude, Gemini, DeepSeek, Grok and Kimi models. Models with reasoning are marked with a reasoning icon in the model selector, Gemini models with reasoning show a thinking summary, and a reasoning effort selector in the chat lets you choose how hard the model thinks.
When should you use a reasoning model?
Use reasoning when the answer depends on getting several steps right in a row, or when a wrong answer is expensive. Good fits:
- Coding and debugging: tracing a bug across files, designing a data model, reviewing a tricky change. If you build with AI, see what vibe coding is and where it needs a careful model.
- Math and logic: word problems, calculations with several stages, proofs, puzzles.
- Planning: project plans, schedules with many constraints, comparing options against criteria.
- Analysis: reading a long document and drawing conclusions, spotting contradictions, checking an argument.
- Agent tasks: multi-step work where the model must choose tools and check results. More in What is an AI agent?
A prompt that suits a reasoning model:
"Here is my function and the error it throws. Find the root cause, explain why it happens, propose the smallest fix, and tell me how to verify it. List any assumptions you made."
When a regular model is the better choice
Reasoning is not free. For most everyday requests it only adds waiting time. Stay with a fast model for:
- Quick facts, definitions, and casual chat.
- Rewriting, translating, and tone changes (see how to make AI writing sound less generic).
- Short summaries and brainstorming lists.
- Image prompts and creative first drafts.
A simple rule: start with a fast model. If the answer is shallow, wrong, or misses a constraint, send the same prompt to a reasoning model, or raise the effort. Comparing both answers side by side is often the quickest way to see whether the extra thinking helped.
What is reasoning effort?
Reasoning effort is a setting that controls how much a reasoning model thinks before answering, usually low, medium, or high. Low effort is quicker and uses fewer tokens; high effort lets the model explore more approaches and check its work more thoroughly.
- Low: simple questions that still benefit from a quick check, like a short calculation.
- Medium: a sensible default for most coding and analysis tasks.
- High: hard bugs, long multi-step problems, and anything you would otherwise double-check by hand.
Thinking also uses the model's context window, so very long chats plus high effort can crowd out your own material. Our guide to AI context windows explains how that budget works.
How to prompt a reasoning model
Prompting a reasoning model is simpler than prompting a regular one, because it plans by itself. A few habits help:
- State the goal and the constraints clearly. What counts as a correct answer, and what must not change.
- Skip "think step by step". The model already does; long scripted instructions can get in the way.
- Give it the real material. The actual code, data, or document beats a description of it.
- Ask for a verification step so you can check the result yourself.
For the basics that apply to every model, see how to prompt AI effectively, and for how the big model families differ by task, read Claude vs ChatGPT vs Gemini.
Try reasoning and non-reasoning models side by side
The fastest way to learn when reasoning pays off is to run your own prompt through both kinds of model. In Chocolatey AI you can switch between models and change the reasoning effort without opening another app. The free plan has a daily message limit, and Pro unlocks premium models (see the FAQ). There is also a desktop app for Windows and macOS.
Frequently asked questions
What is a reasoning model in AI?
A reasoning model is a large language model that works through a problem step by step before it gives its final answer. It spends extra time and tokens on internal "thinking", which makes it slower but usually more accurate on math, coding, planning, and multi-step analysis.
What is the difference between a reasoning and a non-reasoning AI model?
A non-reasoning model starts writing its answer right away, which makes it fast and cheap for chat, drafting, and simple questions. A reasoning model first plans and checks its work, which takes longer and uses more tokens but catches more mistakes on hard, multi-step problems.
Is ChatGPT a reasoning model?
ChatGPT is an app, not a single model. It can run both fast standard models and reasoning models, depending on the model or thinking setting you pick. The same is true of most AI apps that offer several models.
Is Claude a reasoning model?
Recent Claude models can answer normally or use extended thinking before they reply, so they work as both. Whether a given chat uses reasoning depends on the model version and the thinking or effort setting the app gives you.
Is GPT-4o a reasoning model?
No. GPT-4o is a fast general-purpose model that answers directly. OpenAI's reasoning models were released as a separate line, starting with o1 in 2024, and later models combined both modes.
When should I not use a reasoning model?
Skip heavy reasoning for quick questions, casual chat, rewording, short summaries, and brainstorming. Those tasks do not benefit from long thinking, and a fast model returns a good answer in a fraction of the time.