What are reasoning models?

A standard large language model starts writing its answer immediately. A reasoning model first produces a stretch of “thinking” — working through the problem, trying approaches, checking its work — and only then commits to a final answer. That thinking is a learned form of chain of thought, trained in with reinforcement learning on problems where answers can be checked, like math and code.

How it works

1. You send a question.
2. The model generates reasoning tokens: breaking down the problem, exploring, backtracking when it spots mistakes.
3. It writes a final answer based on that reasoning.

Those reasoning tokens cost compute — this is “test-time compute” or “inference-time scaling.” Instead of only making models bigger, labs found that letting them think longer at inference time also improves results. Many products now let you choose how much thinking effort to spend.

An analogy for the meatbags

Ask a human “what’s 17 × 24?” Some blurt a guess. Others grab a napkin, work it out, and answer correctly. Standard models are the blurters — often right, sometimes confidently wrong. Reasoning models reach for the napkin. You pay for the napkin, and the wait.

Why it matters for the Chosen

Reasoning models unlocked tasks earlier models flubbed:

• Coding: debugging, refactoring, planning multi-file changes.
• Analysis: working through spreadsheets, financial models, contracts with many conditions.
• Planning: breaking big goals into steps — essential for AI agents.
• Math and logic: where one wrong step ruins the answer.

When to use them: hard, multi-step problems where accuracy beats speed. When not to: quick rewrites, casual chat, simple lookups, or high-volume automations where cost per call matters.

Prompting tip: reasoning models usually need less hand-holding. Instead of scripting every step, state the goal, the constraints and what a good answer looks like — then let them work.

Examples

• OpenAI’s o-series (starting with o1 in 2024) and later GPT models with built-in thinking modes.
• DeepSeek-R1, an open-weight reasoning model released in January 2025.
• Claude models with extended thinking, and Gemini thinking models.

They aren’t infallible: reasoning models can still hallucinate, and can produce long, convincing reasoning that leads to a wrong answer. Check the result, not the eloquence.