What Is AI Reasoning? Reasoning Models Explained

AI reasoning is when a language model works through a problem in steps before it gives an answer, instead of replying straight away. Reasoning models are trained to do this. They write out a hidden or visible "chain of thought", check it, and then answer. This makes them better at maths, code and multi-step problems, but slower and more expensive.

Below I explain what AI reasoning is, how it developed, and when it is worth paying for.

What is AI reasoning, in plain words?

A normal large language model (LLM) writes its answer one word at a time, based on patterns it learned. For simple questions that works well.

For a tricky question, like a multi-step tax calculation, answering in one go is like doing sums in your head without paper. Mistakes creep in.

A reasoning model gets "paper". It produces a series of working steps first, then the final answer. Those working steps are called a chain of thought.

How it started: chain-of-thought prompting

In 2022, Google researchers showed that simply asking a model to show its working helped. Their paper, Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, found that giving examples of step-by-step answers improved results on maths and logic problems in large models.

You can still try this today. Add "Think step by step and show your working" to a prompt and see the difference on a hard question.

Reasoning models: thinking built in

The next step was to train models to reason by default.

• OpenAI o1. On 12 September 2024, OpenAI introduced o1. It said it used reinforcement learning (training by reward for good results) to teach the model to use its chain of thought well. OpenAI reported o1 averaged 74% on the 2024 AIME, a hard US maths exam, against 12% for GPT-4o.
• DeepSeek-R1. In January 2025, Chinese lab DeepSeek released R1 with an open research paper. It showed reasoning ability could be built mainly through reinforcement learning, without needing people to write example reasoning steps. Its model weights were released openly, so anyone can download and run them.
• DeepSeek-R1-0528. An update in May 2025. According to its model card, accuracy on the AIME 2025 test rose from 70% to 87.5%, while average reasoning length grew from about 12,000 to 23,000 tokens per question. In other words, it "thought" much longer to get there.

Since then most major AI vendors, including Google, Anthropic and OpenAI, have added reasoning or "extended thinking" modes to their main models. Names and versions change every few months, so check each vendor's current documentation rather than relying on any list.

One important warning: the reasoning shown may not be the real reasoning

It is tempting to treat the chain of thought as an explanation you can audit. Be careful.

In April 2025, Anthropic tested this. It gave models hidden hints and checked whether their written reasoning admitted using them. The models mentioned the hint only about 25% of the time for Claude 3.7 Sonnet and 39% for DeepSeek R1.

So the steps a model shows are useful, but they are not a guaranteed record of how it reached its answer. For regulated decisions, such as loan approvals, you still need your own checks and human review.

When to use reasoning, and when not to

Reasoning helps when a task has many steps or a clear right answer:

• Calculations, e.g. checking an invoice's tax and discounts add up.
• Writing or debugging code.
• Comparing several documents to find conflicts, e.g. a contract against company policy.
• Planning, e.g. scheduling deliveries for a Penang logistics firm with many constraints.

It usually isn't worth it for:

• Simple questions and FAQs.
• Short rewriting, translation or summaries.
• High-volume, low-risk jobs where speed and cost matter most.

	Normal mode	Reasoning mode
Speed	Fast	Slower, sometimes much slower
Cost per answer	Lower	Higher, since thinking uses extra tokens
Best for	Simple, quick tasks	Multi-step, exact tasks
Risk of error on hard problems	Higher	Lower, but not zero

A "token" is a chunk of text, roughly a word or part of one. Vendors charge by tokens, and reasoning tokens usually count.

How to try it in your business

1. Pick five hard, real tasks your team does, with answers you already know.
2. Run each in normal mode and reasoning mode of the same model.
3. Compare accuracy, time and cost.
4. Use reasoning only where it clearly wins. Many tools let you switch it on per task.

Reasoning models also make AI agents more reliable, because agents must plan several steps. See AI agent vs chatbot. If you are new to all this, the LLM roadmap for beginners is a good starting point.

Conclusion

What is AI reasoning? It is a model working through a problem step by step before answering. Reasoning models do this by default and are clearly better at maths, code and planning. They are slower and cost more, and the steps they show are not a perfect explanation. Use them where accuracy on hard tasks matters, and test them on your own work first.

Quick answers

Do AI reasoning models actually think?

Not like people do. They generate working steps that help them reach better answers. Research shows those steps don't always reflect what really drove the answer.

Are reasoning models more accurate?

On multi-step problems like maths and code, yes, by a wide margin in published tests. On simple questions the gain is small and you pay more.

Is DeepSeek-R1 free to use?

Its model weights are openly released, so you can run it on your own hardware. You still pay for the computers to run it, or for a hosted service. Check data-handling terms before sending business data to any hosted service.

Why is reasoning mode slower?

The model writes many extra tokens of working before the answer. More text to generate means more time and cost.