🤖 AI / ML

Reasoning Model

thinking model · large reasoning model · LRM

Definition

A reasoning model is an LLM that spends extra computation on a chain of reasoning before its final answer, improving accuracy on multi-step tasks at the cost of latency and price.

How a reasoning model works

A standard LLM generates the answer token by token as soon as it receives the request. A reasoning model first writes a long internal chain of reasoning: it breaks the task into steps, tries options, spots mistakes and backtracks. The user sees only the final answer — in OpenAI o1 the chain of reasoning is hidden by design.

Models learn this behaviour through reinforcement learning. In the DeepSeek-R1 paper, the authors showed that reasoning ability can be developed through “pure” reinforcement learning, without human-labelled reasoning examples. The price is paid in computation at the inference stage: the longer the model “thinks”, the more accurate the answer and the more expensive it is.

Milestones

Date	Event
12 September 2024	OpenAI releases o1-preview and o1-mini; in the announcement the company reports that o1 solves 83% of problems on the AIME maths exam versus 13% for GPT-4o
5 December 2024	The full version of o1 is released
20 January 2025	DeepSeek-R1: open weights, MIT licence, six smaller distilled models
24 February 2025	Claude 3.7 Sonnet — a hybrid model: standard mode and extended thinking in one model

As of September 2026, the depth of reasoning in major APIs is set by a request parameter — reasoning effort at OpenAI, a budget or adaptive thinking mode at Anthropic.

Pros and cons for e-commerce

Assistant task	Standard LLM	Reasoning model
Delivery, returns, order status	Fast and good enough	Overkill: slower and more expensive
Selection by one or two parameters	Copes well	Small gain
Comparing 5 products on 6 criteria	May mix up specifications	A strength
Bundle within a budget, compatibility check	More often gets arithmetic and conditions wrong	A strength
Multi-step agent actions	More often loses the plan	Plans and checks each step

Reasoning does not eliminate hallucination: if the context contains no data about a product, the model can still invent it — just with better arguments.

In practice: routing in an online assistant

Classify the request. A fast classifier or a set of rules decides whether the question is simple or complex.
Simple questions go to a fast model. FAQs, availability, delivery, size checks.
Complex ones go to a reasoning model. Comparisons, bundles, compatibility, multi-step AI agent scenarios.
Set a budget. Cap the depth of reasoning and the response time; for the first reaction, show a short line such as “Comparing the models against your criteria”.
Measure. Compare p95 latency, cost per conversation and conversion in an A/B test, not just answer quality on test questions.

Frequently asked questions

How is a reasoning model different from a standard LLM?

A standard LLM starts writing the answer straight away. A reasoning model first generates a hidden chain of reasoning — trying options, checking itself — and only then gives the result. OpenAI trained o1 to do this with reinforcement learning and noted that accuracy grows with the amount of compute spent “thinking”.

Why is a reasoning model slower and more expensive?

Because it generates more tokens. According to OpenAI's documentation, reasoning tokens are not visible through the API but occupy space in the context window and are billed as output tokens; Anthropic also bills thinking tokens as output. Many APIs let you tune the depth of reasoning: at OpenAI it is the reasoning effort parameter — low effort is faster and cheaper, high effort more thorough.

What is a hybrid reasoning model?

A model that can answer either straight away or after reasoning. On 24 February 2025 Anthropic introduced Claude 3.7 Sonnet as the first hybrid reasoning model on the market: the developer chooses standard mode or extended thinking and can cap the reasoning budget at a set number of tokens.

Does a shopping assistant need a reasoning model?

Not for every question. “How much is delivery?” or “Do you have it in size 42?” need no reasoning, and extra seconds of waiting only irritate the shopper. A reasoning model earns its cost on complex requests: putting together a bundle within a budget, checking compatibility, comparing five laptops on six parameters.