Thinking models

In Model differences and harnesses you met the idea that every AI company ships a fast model and a deeper one. This lesson opens up the deeper one: what it actually does differently, when it is worth the wait, and how to find the switch in the three big chat apps.

What "thinking" actually means

A thinking model (also called a reasoning model) is a model that writes out working notes before it gives you an answer. That is the whole trick. Instead of going straight from your question to a reply, it first produces a stretch of private scratch work: restating the problem, listing what it knows, trying an approach, noticing a mistake, trying again. Only then does it write the answer you see.

Why would that help? Remember that a model builds its reply one piece at a time, and each new piece is shaped by everything already written. When the first thing it writes is the final answer, it has to get the answer right with no room to work. When the first thing it writes is a plan and a few checks, the final answer gets to lean on all of that. It is the difference between blurting out a total and doing the sum on the back of an envelope first.

The scratch work is not magic and it is not a different kind of brain. It is the same sort of text the model always writes. The companies trained these models to get good at writing useful notes for themselves, and to use those notes when they answer. The apps label this in different ways, "Thinking," "Extended thinking," "Think," or a thinking level, but they all describe the same move: work first, answer second.

What it is good at, and what it is not

Thinking pays off when a job has several steps that depend on each other, or when one slip early ruins everything after it. OpenAI's own guidance for builders describes reasoning models as strong at planning, strategizing, and making decisions from a pile of messy information, and at precise work in areas like math, science, and finance. In everyday terms that means:

• Multi-step logic. "If the crew finishes the Washington job by noon, can they make the St. George job, given the drive and the lunch break?"
• Math word problems. Splitting a bill, comparing two loan offers, working out how much mulch a yard needs.
• Planning. A schedule with rules that all have to hold at once, a move, a week of meals around three picky eaters.
• Checking. "Here is my quote, does the total add up?" or "Find the flaw in this plan."

Thinking is a poor fit for jobs with nothing to work out. A simple lookup ("what does PTO stand for?"), a quick rewrite, a friendly back and forth, brainstorming fun names for a food truck. The fast model gets those right on the first try, and the thinking model just makes you wait for the same answer. It is also the wrong pick when speed is the point, like drafting a reply while a customer is still on the phone.

Tap through the three tasks below. Watch which lane wins, and notice that the thinking lane is always slower, even when it loses.

Why it is slower and eats your limit

Those working notes are made of the same tokens as everything else the model writes, and every token takes time and computing power to produce. A fast answer might be a few hundred tokens. A thinking answer can be the same few hundred tokens of reply sitting on top of thousands of tokens of notes you mostly never read. You wait for all of it, and the company pays to run all of it.

That is why chat apps put thinking behind limits. As of October 2026, Anthropic's help pages say higher effort settings in Claude "take longer and use more tokens," so you reach your usage limit faster. Google's Gemini help pages say more advanced models and higher thinking levels "consume more of your usage." OpenAI says that if you hit a Thinking limit in ChatGPT, it may continue with another available model. The exact numbers change often and differ by plan, so treat thinking like the good printer paper: there when you need it, wasted on a grocery list.

Where the switch lives, as of October 2026

These menus move around a lot, so the general rule matters more than the exact clicks: tap the model name near the box where you type, and look for the word thinking, effort, or a level.

• ChatGPT. Open the model picker. Instant gives quick replies, and on eligible paid plans a Thinking slider sets how much it thinks, from Medium up to Extra High. On Plus and Pro, Instant does not move up a level on its own, and typing "think harder" does not change the setting. Free and Go users get a Think button for harder questions.
• Claude. Click the model name next to the send button. Depending on the model, you either flip a Thinking or Extended toggle, or pick an effort level from Low up to Max. On several current Claude models thinking cannot be switched off at all, and the effort level is your only dial.
• Gemini. Tap the model name inside the text box to choose between Flash-Lite, Flash, and Pro. Then look for Thinking Level if your plan shows it: Standard is the default, Extended reasons longer, and Deep Think, the slowest and most thorough, is only for Google AI Ultra subscribers.

Some apps also decide for you. If your app is set to switch automatically, a hard question may quietly get the thinking treatment, and a quick one may not. That is fine for most people. Pick it by hand when you know the job is a planning or checking job.

The thoughts you see are a summary

When thinking is on, most apps show a little timer and an expandable section, often labeled "Thinking," above the answer. It is natural to read that as a live window into the machine's mind. It is not quite that.

Anthropic says the Thinking section in Claude shows a summary of how Claude approached the problem, not its full reasoning. OpenAI made the same call when it launched its first reasoning models: it chose not to show the raw notes and shows a model-written summary instead. So what you read is a second, tidier piece of writing about the notes, not the notes themselves.

Two practical lessons come out of that. First, the summary is useful for spotting a misunderstanding early ("it thinks I meant the Hurricane location, not Washington"), so it is worth a glance. Second, a confident-looking summary does not prove the answer is right. Even the full notes are just more text the model wrote, not a printout of the gears turning inside it. Check the answer the same way you would check any AI answer, as covered in Trust and limits.

Try this yourself

Open whichever chat app you already use. Paste the prompt below twice: once with the fast option, and once with thinking turned on (in free ChatGPT, that is the Think button). You do not need a paid plan for this.

A landscaping crew in St. George has three jobs tomorrow. Job A takes 3 hours and must finish before noon. Job B takes 2 hours and the customer is only home after 1 pm. Job C takes 4 hours and is a 40 minute drive from the other two. The crew starts at 7 am, needs a 30 minute lunch, and must be done by 5 pm. Can they do all three? If yes, give the order and times. If no, say which job to move and why.

Then compare, without judging yet:

• How long did each one take? Did you see a timer?
• Open the thinking section. Does it read like raw scratch work, or like a neat summary someone wrote afterward?
• Do the two answers agree? If not, check the times yourself. Which one held up?
• Now ask both modes something simple, like "what is a good name for a shaved ice stand?" Was the wait worth it this time?

That last comparison is the whole lesson in one minute: thinking for jobs with steps, fast for everything else.