Reasoning models explained: when a pause gives a better answer
Some models work through a problem before they answer. In our chat the Think chip turns that on. It helps a lot with maths, logic and code and very little with a quick rewrite. Here is how to tell.

In the composer of our chat there is a small chip labelled Think. Press it and your next message goes to GPT-6.1 Terra instead of the fast default, GPT-6 Luna. Terra spends a moment working on the problem before the first word of the reply appears, and the chat then shows a short line such as "Thought for 3 s". Models that do this are usually called reasoning models. This guide explains what happens during that pause, which tasks get better because of it, and a simple rule for when to press Think.
What "thinking" means for a language model
A language model produces text one small piece at a time, and each new piece depends on what came before. A model that answers straight away commits to its approach in the first few words. If that approach is wrong, the rest of the answer usually follows it faithfully, because the model is building on its own start.
A reasoning model is trained to write a stretch of private working before the reply. In that working it can restate the problem, list the conditions, try a route, spot a contradiction and back out of it. Only after that does it write the answer you read. The working is real text that the model generates, so it costs time even though you never see it.
OpenAI's documentation describes this with a setting called reasoning effort. Depending on the model, it ranges from none through low, medium and high up to extra high and max. Lower effort favours speed; higher effort lets the model think more fully. OpenAI also notes that the models adjust on their own, spending fewer thinking tokens on simple tasks even at the same setting.
Luna and Terra in our chat
We run both models through a cloud API and choose the effort for you, so the chat stays quick:
| Model | Thinks before answering? | How to choose it | Good for |
|---|---|---|---|
| GPT-6 Luna | As little as possible: lowest effort | Default; Think off | Writing, rewriting, summaries, translation, quick questions |
| GPT-6.1 Terra | Yes, briefly: low effort | Press Think, or pick it in the model menu | Maths, logic, planning, debugging, careful comparisons, detailed images |
| GPT-6 Sol | Not offered here | Locked on gpt.me | Available in our AI Chat app with a Pro subscription and on ChatPlus |
In a live test on a short question, Luna started answering in about one and a half seconds and Terra in under two. That gap grows on harder problems, because Terra thinks longer when the task needs it. The thinking itself stays hidden: the chat shows only how long it took, not the notes. Read more on the GPT-6.1 Terra page and in our Luna vs Terra comparison.
Tasks that deserve the pause
Thinking helps whenever one early slip would ruin the result, or when several conditions must hold at once.
- Calculations with several steps. Loan repayments, recipe scaling, splitting costs unevenly, unit conversions inside a word problem.
- Scheduling and constraints. Shift rotas, tournament draws, a study timetable that has to fit around a part-time job.
- Debugging. Following what a loop does on each pass is exactly the kind of tracing the working is for.
- Decisions with trade-offs. Two flats, two job offers or three phone plans compared on cost, distance and contract terms.
- Finding the flaw. The weak step in a proof, a hidden assumption in a business plan, the logical jump in an essay.
- Dense images. OpenAI describes Terra as strong at detailed visual reasoning, such as reading technical drawings or busy charts.
A task worth pressing Think for: Three friends rent a holiday house for 7 nights at 840 in total. Ana stays all 7 nights, Ben stays 5, Chloe stays 3. They want to split the rent by nights stayed, but Ben already paid 300 upfront. How much does each person still owe, and who pays whom?
Send it to Terra, then send the same question to Luna in a new chat and compare how each one handles Ben's advance payment.
Tasks where Luna is enough
A large share of everyday requests have no hidden steps, and thinking first only adds a wait. Leave Think off for:
- changing the tone or length of a text you paste;
- first drafts of emails, posts, product descriptions and stories;
- explaining a word, an idea or a historical event in simple terms;
- brainstorming, where variety matters more than precision;
- translating ordinary text.
OpenAI's own guidance points the same way: it recommends low effort for routine work such as small edits and fact extraction, and higher effort for hard debugging and deeper analysis.
How to prompt a thinking model
A reasoning model is good at working out how to solve a problem. It cannot work out facts you never gave it. Put every condition into the first message.
Weak: Make a revision timetable for my exams. Strong: Make a revision timetable for the next 14 days. Exams: maths on day 9, chemistry on day 12, history on day 14. I work at a cafe on days 2, 5 and 8 from 10:00 to 16:00. I can study at most 5 hours a day and want one free evening a week. Give a table by day, then list any rule you could not meet.
Some habits that help:
- Skip "think step by step". OpenAI's guidance for reasoning models calls it unnecessary, since they already reason internally. Ask for the steps to be shown in the answer if you want to check them.
- State the goal and the format. "Give the final figure on its own last line" makes the result easy to find.
- Ask for assumptions. When information is missing, a thinking model still guesses. "List anything you assumed" brings those guesses into view.
- Stay sceptical. Careful reasoning on a wrong fact still gives a wrong answer. Our fact-checking guide shows what to verify.
Run your own comparison
Pick a problem you already know the answer to: a solved homework question, a bug you fixed last month, a schedule you worked out by hand. Ask Luna first, then regenerate the answer with Think on, and flip between the two versions with the arrows under the reply. Look at where each one goes wrong, not just at the final line. After a few tries you will know which of your own tasks are worth the extra seconds. The math solver and the code assistant are good places for such tests, and how language models work explains why working first helps.
FAQ
Can I read what Terra was thinking?
No. Our chat asks the model to keep its working private and shows only the time it spent, as in "Thought for 4 s". If you want the reasoning, ask Terra to explain its steps in the answer itself.
Does a Think answer cost more of my free messages?
No. Each answer counts as one message whether Think is on or off. The only cost is the wait. The FAQ explains the daily allowance.
Is Terra always more accurate than Luna?
Not always. It tends to do better on problems with steps, but its knowledge has a cut-off and it can still be wrong. For quick rewrites the two models usually give similar results.
Why is GPT-6 Sol locked?
Sol is a higher tier in OpenAI's GPT-6 family, built for complex coding and agentic work. On gpt.me we offer Luna and Terra; Sol is available in our AI Chat app with a Pro subscription and on ChatPlus. The GPT-6 Sol page lists where it is available, including our app.
Does Think stay on for the whole chat?
It applies to the messages you send while it is on, and you can switch it off at any time. Turn it on for the hard step of a conversation and back off for the easy parts.
Got a question? Ask it now
No account and no download. Open the chat, type what is on your mind, and the answer starts appearing in a second or two.
Start chatting

