Reasoning Model

A reasoning model is an AI model that has been trained or configured to engage in explicit step-by-step thinking before producing a final answer. Instead of generating an immediate response, it works through the problem methodically — much like a person showing their working on an exam.

How reasoning models differ from standard models

Standard language models generate responses token by token, optimising for the most likely next word. They can produce fluent, confident answers that are sometimes wrong because they skip over the logical steps needed to arrive at the correct conclusion.

Reasoning models add an explicit "thinking" phase. Before generating the final answer, the model:

• Breaks the problem into components
• Works through each component systematically
• Checks its own logic for errors
• Considers alternative approaches
• Synthesises its analysis into a final answer

Why reasoning models matter

For straightforward tasks — writing an email, summarising a document, translating text — standard models perform well. But for tasks requiring multi-step logic, mathematical reasoning, complex analysis, or careful interpretation, reasoning models significantly outperform standard models. The improvement is especially dramatic for:

• Mathematics and logic puzzles: Problems with a single correct answer that requires several deduction steps.
• Code generation: Complex programming tasks where the model must consider architecture, edge cases, and dependencies.
• Strategic analysis: Business questions where the model must weigh multiple factors and trade-offs.
• Scientific reasoning: Interpreting data, identifying patterns, and drawing valid conclusions.

Examples of reasoning models

• Claude with extended thinking: Anthropic's approach where Claude can use a dedicated thinking space before responding.
• OpenAI o1/o3 series: Models specifically trained for chain-of-thought reasoning.
• DeepSeek-R1: An open-source reasoning model that demonstrates the approach is not limited to closed-source systems.

Trade-offs

Reasoning models are slower and more expensive than standard models because the thinking phase generates additional tokens. For simple tasks, the extra cost provides no benefit. The skill is knowing when to use a reasoning model versus a standard model — match the tool to the task.

The trajectory

Reasoning capability is improving rapidly and is likely to become a standard feature of all major AI models rather than a separate category.