Thinking Mode: When to Make the AI Reason Before It Answers

How it works

In Thinking mode, the AI works through its reasoning before giving its answer: it lays out the data, runs the intermediate calculations, checks them, then concludes. Example: “Is reselling sneakers actually profitable?” The AI calculates the margin per pair, subtracts fees and shipping, then finds the break-even point (say 42 pairs a month) before giving its verdict.

When to turn it on

• Calculations and finance: profitability, budgets, scenarios.
• Strategy: action plans, choosing between several options.
• Complex code and reverse engineering: hard bugs, algorithms, analyzing a program.
• Logic and math: multi-step problems.

For a simple question, a translation or a short text, Thinking mode adds nothing and costs more.

Thinking or a normal answer: the quick test

The simple rule: if you would grab scratch paper to answer it yourself, turn on Thinking. Otherwise, save your gems.

Which models, and how much extra?

The surcharge is added to the message price, in gems:

What it really costs: concrete numbers

Take the Platinum plan and its 6,000 gems per day. Here is how many messages you can send with and without Thinking:

Two takeaways:

• On most models, the surcharge is about a third of the message price: +40 on 150 for DeepSeek v4 Pro, +180 on 550 for Claude Sonnet 4.6, +250 on 800 for Claude Opus 4.7. On a small model like Qwen 3 235B, it more than doubles the price (+20 on 15), but stays tiny in absolute terms.
• A mid-range model with Thinking often costs far less than a top model without it: DeepSeek v4 Pro with Thinking (190 gems) costs under a quarter of Claude Opus 4.7 in normal mode (800 gems). For a calculation or a logic problem, start there.

The most cost-effective strategy

1. Try without Thinking firstOn a suitable model. If the answer is right, you saved the surcharge.
2. If the answer is wrong or shallowRerun the same question with Thinking on, rather than jumping straight to the most expensive model.
3. Move up a tier lastClaude Sonnet 4.6 or Claude Opus 4.7 with Thinking, for the truly hard problems (architecture, an untraceable bug, long reasoning).
4. Give all the data at onceNumbers, constraints, goal. Reasoning on incomplete data is still wrong, however well it is laid out.

Which Thinking model for which use

For code specifically, Thinking helps most when you need to understand before you write: an intermittent bug, a side effect, an architecture decision. To generate a simple script, a code specialist without Thinking like Codestral (75 gems) is enough. More in our guide to coding with AI.

How to write a question for Thinking mode

Thinking performs best when the question is structured. Example:

“I resell sneakers. Average purchase price: €90, resale price: €130, platform fees: 10%, shipping: €8 per pair, fixed costs: €900 per month. Calculate the margin per pair and the monthly break-even point, then tell me whether 30 sales a month is enough. Check your math before concluding.”

• Give the numbers instead of letting the AI guess them.
• Spell out the expected steps: the AI follows them and you can check each one.
• Ask for a check and a clear conclusion (yes or no, which option).

Read the reasoning, not just the answer

The value of Thinking goes beyond a better answer: you see how the AI gets there. Skim the steps and spot the assumption that doesn’t fit your situation, the number copied wrong or the option it missed. Then rerun with just that point corrected (“shipping is €12, not €8”) instead of rewriting the whole question.

Thinking hard, even about uncomfortable topics

Thinking mode is for questions that need reasoning. The problem: life’s truly hard questions are rarely polite: a suspicious partner, a family conflict, a moral dilemma, a legal risk.

On DarkGPT, DeepSeek v4 Pro, Gemini 2.5 Pro or Grok 4.2 in Unfiltered mode reason step by step on those topics too, without dodging or ending with “talk to a professional”. You get a frank analysis, then you decide.

Adults 18+ only; illegal use remains banned by the terms.

The limits

Thinking reduces reasoning errors; it doesn’t eliminate them. If a starting figure is wrong, so is the conclusion. For an important decision (money, health, legal), stay critical and double-check the key numbers. And a longer answer is not necessarily a better one: for a simple factual question, normal mode is faster and cheaper.

Which plan?

Thinking mode is part of the Platinum plan (€89.99/month), and it is unlimited with Diamond (€189.99/month). See all plans, and our Claude Opus vs Sonnet comparison to choose a top-tier model.

Frequently asked questions