Reasoning model
Also known as: reasoning models, thinking model, large reasoning model, LRM
A reasoning model is a large language model trained to 'think' before it answers, generating an internal chain of reasoning steps, checking its work, and trying different approaches before giving a final response.
This makes it much better at math, coding, science, and multi-step planning. While reasoning models are more accurate on hard problems, they're slower and more expensive per question, and their visible 'thinking' isn't always a faithful account of how they reached an answer.
Reasoning model, explained at three levels
At home
A regular chatbot answers like someone blurting out the first thought that comes to mind. A reasoning model is more like a student working the problem on scrap paper first: it writes out steps, notices mistakes, tries another approach, and only then gives its answer. That scrap-paper work takes extra seconds, but it makes the model far better at math and logic.
At work
A reasoning model is a language model trained to work through a problem in steps before answering. It is much stronger at analysis, coding, financial calculations, and planning, where a quick guess is often wrong. For a business, the trade-off is time and money: answers can take seconds to minutes, and the hidden working is billed, so it suits tasks where accuracy matters most.
At home
When you see a 'thinking' indicator, the model is generating a long chain of text you mostly don't see, sometimes thousands of words. In it, the model breaks the question into parts, tests ideas, and checks its own work, much like you might talk through a problem out loud. Each step becomes part of what it reads next, so later steps can catch errors in earlier ones.
At work
These models are trained with reinforcement learning on problems whose answers can be checked automatically, such as math solutions and code that must pass tests. Reasoning paths that reach correct answers are rewarded, so the model learns to plan, verify, and backtrack. APIs from OpenAI, Anthropic, and Google expose a 'reasoning effort' setting or thinking-token budget. Thinking tokens are billed as output, so costs can be several times higher than a standard response.
At home
The model's visible reasoning is generated text, not a recording of its internal process. Research from Anthropic and others found models sometimes reach an answer for one reason and write out a different justification. While the steps genuinely help the model think, they can make a wrong answer sound carefully worked out. In my opinion, it's still wise to check important answers yourself, especially the final numbers.
At work
Reasoning shifts compute from training to inference, so cost and latency scale with how long the model thinks. Production systems route simple requests to fast models, send only hard ones to reasoning modes, set thinking budgets, and cap timeouts. Benchmarks such as AIME and SWE-bench show big gains, but I'd evaluate on your own tasks. Since chain-of-thought may not faithfully reflect the model's actual process, it is useful for safety monitoring yet weak as an audit trail.
Real-world examples
At home
Nadia's son used a reasoning model to check his calculus homework. It took about 30 seconds to answer, but it explained each step and found the error in his solution.
At work
Thomas is a data analyst at a Regina utility. He uses a reasoning model for complicated rate calculations and a faster standard model for routine email drafts, keeping costs in check.
How it compares
| feature | Reasoning model | Standard language model |
|---|---|---|
| Response time | Seconds to minutes | Near-instant |
| Hard math and code | Strong | Weaker |
| Cost per answer | Higher | Lower |
| Best for | Complex, multi-step problems | Simple questions, writing, chat |
Reasoning models trade speed and cost for accuracy on hard problems.
For analysts
I'd use reasoning models selectively, for tasks where mistakes are expensive. Routing everyday requests to faster models keeps both costs and wait times down.
History & origin
Researchers at Google showed in 2022 that prompting models to write out steps, known as chain-of-thought prompting, improved reasoning. OpenAI released o1-preview in September 2024, the first widely used model trained with reinforcement learning to reason at length before answering. In January 2025, China's DeepSeek released R1, an open-weight reasoning model that rivalled o1 at much lower reported cost, contributing to a one-day drop of roughly 17 percent in Nvidia's share price. Anthropic, Google, and others added 'extended thinking' modes to their models through 2025, and many models now switch between fast and deliberate modes automatically. In my view, reasoning models marked the shift from AI that sounds right to AI that more often is right, though not always.