DeepSeek R1 Explained: How the Reasoning Model Thinks, a Worked Example, and How to Access It

What is DeepSeek R1?

Most chat models produce an answer in one pass: they read your message and start generating the reply. A reasoning model adds a stage before that. R1 first writes a private "thinking" passage - restating the problem, planning an approach, trying it, checking the result - and only then writes the answer you see. In the DeepSeek app this thinking is shown in a collapsible panel, so you can audit how it got there.

R1 was trained with reinforcement learning that rewarded correct final answers on verifiable problems (maths, code, logic). Rather than being shown how to reason, the model learned on its own that spending more tokens planning, verifying and backtracking improves its hit rate - which is why its reasoning reads like someone talking themselves through a puzzle. DeepSeek released R1's weights openly, along with smaller distilled versions, which made it one of the most widely used and studied reasoning models available.

How R1's reasoning works

Here's the loop, simplified. Every R1 answer goes through some version of it, and on hard problems it may cycle several times.

1. Understand. R1 restates the question in its own words and lists the constraints - a surprising number of errors in ordinary models come from skipping this.
2. Plan. It picks a strategy ("work backwards", "set up an equation", "enumerate the cases") and sometimes notes alternatives.
3. Work. It executes the steps explicitly, writing intermediate results down rather than holding them in its head.
4. Verify. It checks the result against the constraints, re-does arithmetic, or tests edge cases. If something fails, it says so - "wait, that's wrong" - and returns to step 2.
5. Answer. Only now does it write the reply you'd normally see, usually short and confident because the work is already done.

A worked example: watch R1 solve a problem

Here's a classic word problem that one-pass models often get wrong because the obvious first answer is a trap. The thinking panel is lightly trimmed for length; the structure is exactly what R1 produces.

What to notice

• The verify step explicitly tests the intuitive wrong answer - this is the behaviour that makes R1 reliable on trick questions.
• Intermediate values are written down, so a slip in arithmetic would be visible and easy to point out.
• The final answer is short. All the length lives in the thinking, which you can collapse.

Try it yourself: open the chat, switch on DeepThink and paste the prompt. Then ask "what if the bat cost Rs 900 more?" and watch it reuse the same plan. Full walkthrough of the toggle in How to use DeepThink.

When to use R1 - and when not to

R1 vs V4.1 Flash vs V4 Pro

Wider comparison of all four models, including Coder: DeepSeek models explained · V4 Pro vs V4.1 Flash · DeepSeek vs ChatGPT.

How to access DeepSeek R1

Four routes, from zero-setup to full control. Pick the one that matches how much you want to manage yourself.

# Python · pip install openai
from openai import OpenAI
client = OpenAI(api_key="YOUR_KEY", base_url="https://api.deepseek.com")
r = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=[{"role": "user", "content": "A bat and ball cost 1100…"}],
)
print(r.choices[0].message.reasoning_content)  # the thinking
print(r.choices[0].message.content)            # the answer

Prompting tips for R1

R1 behaves a little differently from ordinary chat models. Here's what to do, why it helps, and what it looks like in practice:

Limits and cost

R1 is powerful but not free of trade-offs. Here's each limit in plain terms and the simplest way around it:

DeepSeek R1 - frequently asked questions