Reasoning Models Explained

Since late 2024, a new type of AI model has appeared: reasoning models. They tackle hard problems far better than standard models — but they work differently and are not always the right choice. This page explains what they do and when to use them.

What Makes a Reasoning Model Different

A standard model (like GPT-4o or Claude Sonnet) reads your message and immediately starts generating a response. It's fast but can miss nuance in complex problems.

A reasoning model pauses first. Before giving you an answer, it generates a long internal "thinking" trace — working through the problem step by step, trying different approaches, checking its own logic, and backtracking when it spots a mistake. Only after that thinking process does it produce a final answer.

Why This Makes Models Much Better at Hard Problems

The improvement on difficult tasks is dramatic. On AIME 2024 (a hard maths olympiad exam), DeepSeek-R1 went from 15.6% to 77.9% accuracy through reinforcement learning that rewarded correct reasoning. OpenAI's o3 scores above 90% on maths benchmarks where standard GPT-4 struggles to reach 50%.

The reasoning process helps because:

• Decomposition — Breaking a hard problem into steps makes each step easier to get right
• Self-checking — The model can catch its own mistakes mid-thought and correct them before the final answer
• Backtracking — If an approach fails, the model can try a different path
• More compute on hard parts — The model spends more tokens on the difficult parts of a problem rather than rushing through them

The Major Reasoning Models (2025)

The Trade-off: Better Answers, But Slower and Costlier

Reasoning models are not always better. The thinking tokens cost money and time. For a simple question like "What is the capital of France?", a reasoning model is overkill — it wastes tokens thinking about something that doesn't need it.

Adjustable Thinking: Fast vs Thorough

Some reasoning models let you control how much they think. This is sometimes called a "thinking budget" or "reasoning effort level."

• Low effort — quicker, cheaper, good for moderately hard problems
• High effort — longer thinking, better answers on the hardest problems, but noticeably slower and more expensive

o4-mini on low effort is often a good sweet spot — much smarter than a standard model at a fraction of o3's cost. Start there and upgrade only if needed.

DeepSeek-R1: Open-Weight Reasoning

When DeepSeek-R1 launched in January 2025, it shocked the AI world: it matched frontier reasoning model performance while being open-weight — meaning you can download and run it yourself for free.

Key facts about DeepSeek-R1:

• Free to use via DeepSeek's website and API (very cheap API pricing)
• Can be downloaded and run locally via Ollama on a capable home GPU
• Shows its full reasoning trace (not hidden) — good for learning and debugging
• Strong on maths and coding; somewhat weaker on long-form English writing
• Context: trained in China, so data privacy and censorship considerations apply for sensitive topics

A Simple Test: Does Thinking Help?

Not sure whether to use a reasoning model? Try this:

1. Ask a standard model (e.g. GPT-4o or Claude Sonnet) your question
2. If the answer looks right and you can verify it — done
3. If the answer seems wrong, inconsistent, or incomplete — try o3, o4-mini, or DeepSeek-R1
4. If the reasoning model also struggles, the problem may need a different approach (more context, clearer framing, or a specialist tool)

You'll quickly develop an intuition for which types of problems benefit. Anything that involves counting steps, tracking constraints, or doing calculations is a strong candidate.

Checklist: Do You Understand This?

• Can you explain in plain words what a reasoning model does differently from a standard model?
• Can you name two reasoning models and who makes them?
• Name three tasks where a reasoning model is the better choice.
• Name three tasks where a standard model is fine and the reasoning model is overkill.
• What does "adjustable thinking budget" mean and why does it matter?
• What is special about DeepSeek-R1 compared to other reasoning models?