Reasoning models explained: when it pays to let Grok think

Pick Grok 4.3 Thinking in our chat and you will notice a pause before the first word appears. During that pause the model is not stuck. It is writing itself notes: breaking the question into parts, trying an approach, checking it, sometimes starting over. Models that do this are called reasoning models or thinking models. This guide explains what they do, what the extra seconds buy you, and a simple way to decide between the fast default, Grok 4.20, and Grok 4.3 Thinking for each task.

What a thinking model does differently

Every language model writes its answer one small piece of text at a time, each piece chosen based on everything before it. A regular model goes straight to the answer. If it starts down a wrong path in the first sentence, it tends to stay on it, because later words are built on earlier ones.

A reasoning model has been trained to produce a stretch of working first, a bit like the scrap paper you use before writing the final answer on an exam. In that working it can list the conditions of a problem, try a calculation, notice a contradiction and fix it. Only then does it write the reply you see. The working itself is extra text the model generates, so it takes time and computing power, even when you never see it.

That is the whole trick, and it explains both sides of the trade. Problems that benefit from steps, checks and second tries get better. Problems that need no steps just get slower.

Why you wait, and what you do not see

In our chat the thinking happens out of view. You see a "Thinking…" indicator while Grok 4.3 works, then the final answer streams in. The hidden notes are not shown and are not saved in your history.

xAI's developer documentation describes a setting called reasoning effort, which controls how much thinking a model does before it answers. For Grok 4.3 the options range from none to high, and the default is low. Our chat runs Grok 4.3 at low effort: enough thinking to catch many slips, without making you wait a minute for a reply. Grok 4.7 also reasons before every answer, since xAI's documentation says its reasoning cannot be switched off; we run it at low effort too. Grok 4.20 runs with reasoning switched off, which is why it starts answering almost at once.

On a short question in our own tests, Grok 4.20 began replying in about a second, while Grok 4.3 Thinking took a few seconds. On harder problems the thinking stretch grows, so expect longer pauses.

Tasks where thinking pays off

• Multi-step maths and word problems. Percentages of percentages, unit conversions, probability, anything where one early slip ruins the result.
• Logic and constraints. Seating plans, timetables, "who owns the fish" puzzles, scheduling five people with conflicting availability.
• Debugging. Tracing why a function returns the wrong value means following the code step by step, which is exactly what the working is for.
• Comparisons with trade-offs. Choosing between two job offers, two mortgages or three software tools when several factors pull in different directions.
• Checking an argument. Finding the gap in a proof, the hidden assumption in a business case or the step that does not follow in an essay.

Example worth the wait:
A shop raises a price by 20 percent, then puts it on sale at 25
percent off the new price. A customer pays 54 after the sale. What was the original
price, and is the sale price higher or lower than the original?
Show the steps.

A fast model may produce the right number, but this is the kind of question where it is most likely to get one step wrong while sounding sure. Open Grok 4.3 Thinking and try it, or send the price puzzle to whichever model you have selected and compare.

Tasks where it adds little

For a large share of everyday requests, thinking first only adds a delay. Stay on Grok 4.20 for:

• rewriting, shortening or changing the tone of a text you paste;
• drafting emails, posts, captions and short stories;
• simple facts and definitions (which still need checking, thinking or not);
• brainstorming names, ideas or headlines, where variety matters more than rigour;
• translation of ordinary text.

xAI's own developer cookbook makes a similar point: in one of its examples, higher reasoning effort on short support tickets used more tokens without reliably scoring better. More thinking is not automatically better thinking.

The three models in our chat, side by side

All three read images and all three share the same daily allowance of 30 free messages, so a thinking answer costs you no more than a fast one. The only price is the wait. The Grok 4.3 page and the comparison of Grok 4.7 and Grok 4.3 have more detail.

How to prompt a thinking model

Reasoning models are good at working out how to solve a problem. What they cannot work out is information you left out. Put every condition in the first message.

Weak:
Help me plan meetings for my team next week.

Strong:
Plan five 30-minute one-to-one meetings next week, Monday to Friday,
between 9:00 and 17:00. Rules: I am busy Tuesday morning and every
day 12:00-13:00. Ana only works Monday to Wednesday. Ben cannot do
mornings. No more than two meetings per day. Give a table, then
point out any rule you could not satisfy.

A few more habits help:

• Skip "think step by step". The model already does that before answering. Ask instead for the steps to be shown in the reply if you want to check them.
• Ask for a final line with the answer alone, so it is easy to find.
• Ask it to list any assumptions. Thinking models still fill gaps with guesses.
• Do not treat the extra care as a guarantee. A reasoning model can reason its way neatly to a wrong answer, especially about facts it never learned.

A two-minute experiment

The best way to get a feel for the difference is to see it. Pick a problem you already know the answer to, such as a homework question with a solution or a bug you fixed last week. Send it to Grok 4.20, then start a new chat with Grok 4.3 Thinking and send the same message. Compare not just the final answers but how each one handles the tricky step. After a few rounds you will know which of your own tasks deserve the pause. The math solver and the code assistant are good places to run such tests, and our guide to how language models work explains why the step-by-step approach helps.

FAQ

Can I read Grok's thinking in your chat?

No. Our chat hides the working and shows only the final answer. If you want to see how it got there, ask it to explain its steps in the reply.

Does the thinking model use more of my daily messages?

No. Every answer counts as one message whichever model you pick. Grok 4.7 has its own cap of 10 a day within the 30.

Is Grok 4.3 Thinking always more accurate?

Not always. It usually does better on problems with steps, but it can still be wrong, and it knows no more facts than its training gave it. Check anything important, as our fact-checking guide shows.

Why is Grok 4.20 not a thinking model here?

Grok 4.20 can reason, but we run it with reasoning switched off so the default stays fast for everyday tasks. When you want the careful mode, choose Grok 4.3 Thinking from the model menu.

What is the difference between Grok 4.3 Thinking and Grok 4.7?

Both think before answering. Grok 4.7 is xAI's newest flagship, described for coding, agentic and knowledge work, and it is limited to 10 messages a day in our chat. Grok 4.3 is the everyday choice for careful reasoning.