// how to

How to Use DeepThink in DeepSeek (and When to Leave It Off)

Click DeepThink in the DeepSeek message box before you send your question. DeepSeek then works through the problem step by step and shows that reasoning before its final answer. Use it for math, logic puzzles, code and decisions with several factors; leave it off for quick rewrites, emails and simple questions.

By Published Last updated 2 min read

The short version

DeepThink makes DeepSeek think before it answers. You’ll see the reasoning first, then the answer. It’s slower, and worth it when the problem has steps that can go wrong.

Step by step

  1. Open a chat at chat.deepseek.com or in the app.
  2. Click DeepThink in the message box so it’s highlighted.
  3. Ask the whole problem in one message: the numbers, the rules and what you want at the end.
  4. Watch the reasoning or skip straight to the answer. If the reasoning shows it misread your question, send a clearer version rather than arguing with a wrong answer.
  5. Switch it off when you move on to quick tasks.

When DeepThink helps

  • Math and word problems where one wrong step ruins the result. See solve math problems with DeepSeek.
  • Code: finding a bug, choosing between two approaches, explaining why something fails.
  • Decisions with trade-offs: comparing three quotes, planning a schedule with constraints.
  • Logic and puzzles: scheduling, seating plans, anything with “if this, then not that”.

When to leave it off

  • Rewriting, shortening or translating text
  • Emails, captions and short summaries
  • Simple facts or definitions
  • Brainstorming lists of ideas

For these, the reasoning adds wait time without improving the result.

Prompts to copy

A decision with constraints:

text
I need to choose between three suppliers. Quotes: A $4,200 with 30-day payment terms,
B $3,900 paid upfront, C $4,050 with a 2% early-payment discount within 10 days.
Our cash is tight this month. Compare them and recommend one. Show the math.

A schedule:

text
Plan shifts for 4 staff over 6 days. Each works 4 days, nobody works 3 days in a row,
and Sara can't work Tuesday. Give me a table, then check it against every rule.

Checking its own work:

text
Now go back through your answer and check each step. List anything you got wrong.

Example

You ask (DeepThink on): “Is it cheaper to lease a $40,000 car at $450 a month for 36 months with $3,000 down, or finance it at 6% for 60 months with $5,000 down and sell it after 3 years for $24,000?”

A good result: reasoning that works out each total, notices the resale value matters, and states its assumptions (taxes, fees). Then a short answer you can check with a calculator.

Mistakes to avoid

  • Turning it on for everything. You’ll wait longer for answers that didn’t need it.
  • Splitting the problem across many messages. Give all the facts up front.
  • Treating the reasoning as proof. It can sound careful and still be wrong.
  • Pasting private data to get a “smarter” answer. DeepThink doesn’t change where your data goes. See keep your data private in DeepSeek.

Good to know

In DeepSeek’s API, thinking mode is on by default with “high” effort, and you can set the effort to low, high or max. DeepSeek’s April 24, 2026 announcement also mentioned an Expert Mode and an Instant Mode in the chat. Menus change, so look for the current labels in the message box.

Related: use DeepSeek Search and write better prompts for DeepSeek.

Frequently asked questions

What is DeepThink?

DeepSeek's thinking mode in the chat app. DeepSeek's API docs describe thinking mode this way: before the final answer, the model first outputs a chain of reasoning to improve the accuracy of the response.

Should I trust an answer more because I can see the reasoning?

Not automatically. The reasoning can contain the same kind of mistakes as the answer. Use it to spot where the model went wrong, and check the facts and numbers that matter.

Does DeepThink cost extra?

Not in the chat app, which DeepSeek says is free. In the API, thinking mode is on by default, so turn it off for simple requests if you want shorter, faster replies.

Sources

  1. DeepSeek API Docs: thinking mode
  2. DeepSeek home page (DeepThink and Search switches)
  3. Hugging Face: DeepSeek-R1 model card