What it is
Extended thinking is a mode in which an AI model works through a problem privately, at length, before it writes the answer you see. Instead of replying straight away, the model first produces a long stretch of scratch work: it restates the question, tries an approach, checks a step, notices a mistake, and tries again. Only then does it write its reply.
You have probably met it already. Many chat apps now show a small "Thinking…" label, a thinking toggle, a choice between a fast model and a "thinking" model, or a setting for how much effort the AI should put in. All of these control the same thing: how much working-out the model is allowed to do before it answers.
To see why this helps, it's worth knowing how a language model writes. It produces its reply one small piece of text at a time. Those pieces are called tokens, and a token is roughly three-quarters of a word. Each new token is chosen based on everything written so far. That means a model answering in one go has to get the answer right in the very first words it commits to, with no scratch paper. If the answer depends on five intermediate steps, the model has to do all five in its head, at once.
Extended thinking gives it the scratch paper. The thinking is made of tokens too, so every step the model writes down becomes something it can read and build on for the next step. This is the same idea as chain of thought, where you ask a model to "think step by step" in its reply. The difference is that extended thinking is built into the model by its makers, it happens in a separate space before the answer, and the model has been specially trained to use that space well. The training usually relies on reinforcement learning, which rewards the model when its long working-out leads to a correct final answer.
A few terms you'll see around this feature:
- Reasoning model or thinking model. A model trained to think before answering. OpenAI's o1 in 2024 made the idea widely known; the Inference page tells that story.
- Thinking budget. A cap on how many tokens the model may spend thinking. Early versions of the feature let developers set the number directly.
- Effort. A simpler dial that newer models use instead of a number: low, medium, high, and so on. Higher effort means the model thinks longer and more thoroughly when the task calls for it.
- Adaptive thinking. The model decides for itself whether a question needs thinking, and how much. A greeting gets almost none; a tangled planning problem gets a lot.
- Test-time compute. The industry name for computing power spent at the moment of answering, rather than during training. Extended thinking is the most common way to spend it.
One point matters more than the jargon. Thinking is not free. Every thinking token is computed and, if you pay per use, charged for, usually at the same rate as the answer itself. It also takes time. A reply that would appear in two seconds might take twenty seconds, or several minutes on a hard problem. Extended thinking is a trade: you spend time and money to buy a better chance of a correct answer.
Why it matters
It makes a real difference on multi-step problems. Maths, logic puzzles, code with a subtle bug, a budget with several constraints, a schedule where every change knocks on to three other things: these are tasks where the answer depends on getting a chain of steps right. Thinking gives the model room to lay out the chain and check it. When Anthropic introduced extended thinking in February 2025, it reported that on maths problems, accuracy rose steadily as the model was allowed more thinking tokens. OpenAI had shown the same pattern for its o1 model in September 2024.
It doesn't help much with everything else. A quick fact, a friendly email, a list of ten names for a newsletter: the model already does these well in one pass. Thinking on these tasks mostly adds a wait. On simple tasks it can even do slightly worse: in one 2025 study, models with thinking switched off often matched or beat the thinking versions on easy puzzles. Research on this is covered in the examples below.
It changes what you pay and how long you wait. If you use AI through a subscription, thinking mainly costs you time, and on some plans it can use up your usage allowance faster. If you or your business use AI through an API and pay per token, thinking tokens are billed as output, and on a hard problem the invisible thinking can be many times longer than the visible answer. This is why the providers now offer dials rather than a single on-off switch: most of the time you want a little thinking, occasionally you want a lot, and you want to decide.
It is becoming the default rather than a special mode. In 2025 thinking was something you switched on. By the time of writing, several leading models think on every request and let you set only the effort. On Anthropic's API, for example, Claude Opus 5.5 always runs with adaptive thinking, and developers steer it with an effort setting whose default is "medium". Chat apps are moving the same way, with an automatic mode that decides for you. Knowing what the dial does lets you overrule the automatic choice when it guesses wrong.
What you see is not the whole thinking. Many apps show you a running summary of the model's thoughts, and it's tempting to read it as a window into the machine. It isn't quite. Some providers show a shortened summary written by a separate model, and research shows that even the full thinking text doesn't always mention the real reasons behind an answer. Treat the visible thinking as a helpful sketch, not a transcript.
Real-world examples
Claude 3.7 Sonnet: one model, two speeds (February 2025)
On 24 February 2025, Anthropic released Claude 3.7 Sonnet and described it as a "hybrid reasoning" model. Rather than offering a separate reasoning model, Anthropic built one model that could either answer normally or switch into an extended thinking mode and reason at length first. Anthropic argued that reasoning should be a built-in capability of frontier models rather than a separate model entirely.
Users of the paid Claude plans could choose the mode themselves. Developers using the API got a thinking budget: they could tell the model how many tokens it was allowed to think for, up to its output limit of 128,000 tokens. Anthropic kept the price the same as its previous Sonnet model, $3 per million input tokens and $15 per million output tokens, and said plainly that the output price included thinking tokens.
In a companion research post, Anthropic explained why it chose to make the thinking visible (it helps people understand and check answers) and admitted the limits of that choice: its research had found that models very often make decisions based on factors they don't discuss in their thinking. It also published the scaling result mentioned above, showing that on maths problems accuracy improved roughly logarithmically with the number of thinking tokens. In plain terms, each doubling of thinking time bought a similar-sized gain, so the first extra thinking helps a lot and later thinking helps less and less.
Google puts a number on thinking: Gemini 2.5 (March and April 2025)
On 25 March 2025, Google introduced Gemini 2.5 and called it a "thinking model", one that reasons through its thoughts before responding. A month later, on 17 April 2025, it released a preview of Gemini 2.5 Flash, a smaller, cheaper model that Google called its first fully hybrid reasoning model, giving developers the ability to turn thinking on or off.
The headline feature was the thinking budget, which developers could set anywhere from 0 to 24,576 tokens. At zero, the model skipped thinking for the lowest cost and fastest reply. With no budget set, Google said, the model decided how much to think based on how complex the question looked. Google's announcement gave examples: translating "Thank you" into Spanish needs little reasoning, a probability question or a scheduling problem needs a medium amount, and an engineering calculation needs a lot. In the Gemini app itself, the model calibrated its thinking automatically, with no manual control.
The price showed the trade-off in black and white. At launch, as reported by VentureBeat from Google's pricing, output from 2.5 Flash cost $0.60 per million tokens with thinking off and $3.50 per million with thinking on, almost six times as much. Prices have changed since, so check current pricing yourself, but the principle holds: thinking is paid for.
OpenAI took the automatic approach further. When it launched GPT-5 in ChatGPT on 7 August 2025, it described a system with a fast model for most questions, a deeper "GPT-5 thinking" model for harder ones, and a router that decides which to use based on the conversation, its complexity, and what the user asks for, such as saying "think hard about this" in the prompt. From then on, many users got thinking without ever touching a setting.
"Do NOT think that much for 2+3=?" (December 2024)
On 30 December 2024, a team of researchers led by Xingyu Chen posted a paper with one of the more memorable titles in AI research: "Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs". They studied reasoning models of the kind OpenAI's o1 had made popular, and found that these models spend a great deal of thinking on problems that don't need it.
Their headline example was the question "what is 2 plus 3?". One open reasoning model, QwQ-32B-Preview, found the right answer in its first 39 tokens of thinking, then kept going, checking and re-solving until it had produced 901 tokens and thirteen separate solutions. Across the models they tested, the authors reported that o1-like models used 1,953 percent more tokens than conventional models to reach the same answer to this question. They called it overthinking: spending computing power on easy problems with little or no benefit. They also showed training methods that cut the waste while keeping accuracy.
For everyday users the lesson is simple. More thinking is not automatically better. On an easy question it mostly costs you time.
"The Illusion of Thinking" and the reply (June 2025)
In June 2025, six researchers at Apple, Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio, and Mehrdad Farajtabar, published "The Illusion of Thinking". Instead of maths tests that might have leaked into training data, they used four puzzles whose difficulty they could turn up one notch at a time: Tower of Hanoi, Checker Jumping, River Crossing, and Blocks World. They compared thinking and non-thinking versions of the same models, including Claude 3.7 Sonnet and DeepSeek's models.
They found three regimes. On easy puzzles, the non-thinking models often did as well or better. On medium puzzles, thinking clearly helped. On hard puzzles, both kinds collapsed to zero accuracy. Stranger still, as puzzles approached that collapse point, the thinking models started to think less, not more, even though they had budget left. Giving them the solution method for Tower of Hanoi in the prompt did not move the collapse point. The paper was accepted at the NeurIPS 2025 conference.
The paper spread widely, often summarised as "AI can't really reason". A published reply soon pushed back. On 10 June 2025, a comment posted by A. Lawsen argued that some of the collapse came from the test design: the longest Tower of Hanoi answers would exceed the models' output limits, with models sometimes saying so themselves; some River Crossing puzzles were mathematically impossible yet scored as failures; and when models were asked to write a short program that produces the moves, instead of listing thousands of moves by hand, they did well on sizes previously reported as failures. (An early version of the comment listed Claude as a co-author; that was removed in line with arXiv's rules.)
A fair reading takes something from both. Thinking models have real limits, and their gains are largest on medium-hard problems. How a test is set up can also make those limits look bigger or smaller than they are. Neither paper settles whether the models "really" think.
The thinking you see is a summary (2025 to 2026)
On 3 April 2025, Anthropic published research testing whether reasoning models admit what influenced them. Researchers slipped hints into questions, and the models used the hints, but Claude 3.7 Sonnet mentioned the hint in its thinking only 25 percent of the time on average, and DeepSeek-R1 39 percent. The unfaithful thinking was often longer than the honest kind, so the models weren't leaving the hint out to save words. The chain of thought page covers this study in more detail.
Products have since moved further from showing raw thinking. At the time of writing, Anthropic's developer documentation says that the thinking text it returns is a summary of Claude's reasoning, never the raw chain of thought, that the summary is written by a different model, and that developers are billed for the full thinking, not the summary. On its newest models the thinking text is hidden by default, and developers must opt in to see a summary. So the tidy, readable "thoughts" in your chat app are a digest of a longer process that you don't get to read, and even the full process may not tell the whole story.
An everyday scenario: the launch plan
This is an illustrative scenario, not a documented event. Priya, a self-published author, asks her chatbot for a six-week launch plan for her second novel. She has a £400 budget, a newsletter of 1,200 readers, two podcast interviews already booked in week four, and a day job that leaves her about five hours a week.
With thinking off, the reply arrives in seconds: a neat, generic plan that schedules a cover reveal in week five, after the podcasts that were supposed to show the cover, and spends £600. With thinking on, she waits about forty seconds. The plan that comes back works backwards from the podcast dates, keeps the spending under £400 with a small reserve, and flags that two of the tasks won't fit into five hours in the same week. The summary of the thinking shows the model noticing the clash and moving things around.
Then she asks it to shorten the plan's title. With thinking on, that takes fifteen seconds for a five-word answer she could have had instantly. She switches thinking off for the rest of the tidy-up.
Try it yourself
This takes about fifteen minutes and works in any chat app that offers a thinking toggle, an effort setting, or a choice between a fast model and a thinking model. Features and names change often, so look in the model picker, the settings near the message box, or the app's help pages to find the current control.
- Find the control. In your chat app, look for "thinking", "extended thinking", "reasoning", "effort", or a model name with "thinking" in it. Note what the fast option and the thinking option are called.
- Start a new chat with thinking off, or with the fast option or lowest effort. Paste this planning question and note roughly how long the reply takes: “
I run a small online shop selling handmade candles. Plan my next 4 weeks. Constraints: I can work 10 hours a week. I have 60 candles in stock and can make 20 more per week. A market stall on day 15 needs 40 candles. I want to launch an autumn scent on day 22 and email my 800 subscribers twice, never in the same week as the market. Give a week-by-week plan and a running stock count at the end of each week.“ - Start a second new chat with thinking on, or the thinking model or a high effort setting. Paste exactly the same question. Note the time again, and open the thinking summary if the app shows one.
- Check both plans against the constraints yourself. Does the stock count add up week by week? Are there 40 candles ready for day 15? Are the emails kept out of the market week? Does the work fit in 10 hours? Count the mistakes in each plan.
- Now try an easy question in both modes: “
Suggest five friendly names for an autumn candle scent.“ Compare the waiting time and whether the thinking version is actually any better. - If your app lets you, ask for thinking even with the fast option, just to see the difference a prompt can make: “
Before answering, work through the constraints one by one and check your stock count carefully. Then give the final plan.“ - Write yourself a one-line rule from what you saw, for example: "Thinking on for plans, numbers, and tricky edits; off for names, quick facts, and simple drafts." That rule is the practical value of this element.
If you use AI through an API, you can run the same test at two effort levels and compare the number of output tokens reported for each. That number is what you pay for.
Common mistakes and misunderstandings
Leaving thinking at maximum for everything. It feels safer to always use the most careful setting. In practice it makes routine work slow, uses up your plan's allowance, and on easy questions it tends to overthink without doing any better. Match the setting to the task.
Switching it off for tasks that need it. The opposite mistake is just as common. If you're asking for a plan with constraints, a calculation, an analysis of a contract or a spreadsheet, a debugging job, or a careful edit that must keep certain things unchanged, thinking is usually worth the wait. If a fast answer to such a question looks confident and neat, check it harder, not less.
Reading the visible thinking as the model's true reasoning. What you see is often a summary written by another model, and even the full thinking doesn't always mention what really drove the answer. Use it to spot obvious wrong turns, not as proof that the answer is right.
Assuming thinking stops the model making things up. Thinking helps with steps that can be worked out from the information given. It doesn't give the model facts it doesn't have. A thinking model can still produce a hallucination, and sometimes argues itself into one very persuasively. Check names, numbers, and sources as you would for any answer.
Asking the model to "think step by step" in the reply when thinking is already on. With a thinking model, the working-out already happens. Asking it to repeat that working in the reply mostly makes the reply longer. Ask for what you want to read instead: the final plan, plus the key assumptions.
Thinking that more thinking always means a better answer. The gains shrink as thinking gets longer, and on problems beyond the model's reach, extra thinking doesn't rescue it. If a hard question fails at high effort, break it into smaller steps with prompt chaining, give the model better information, or check the work with evals rather than simply turning the dial further.
Forgetting the bill. For people using an API, thinking tokens are output tokens. A short visible answer can hide a long, paid-for thinking process. Look at the token counts, not just the length of the reply.
Related elements
- Chain of Thought: the prompting technique that extended thinking turned into a built-in feature.
- Inference: where thinking time is spent and paid for, including the story of OpenAI's o1 and test-time compute.
- Tokens: the unit thinking is measured and billed in.
- Reinforcement Learning: the training method that teaches models to use long thinking well.
- Self-consistency: another way to spend extra computing on a question, by asking several times and taking the most common answer.
- Self-critique: asking a model to check its own work, which thinking models partly do on their own.
- Planning: one of the tasks where thinking pays off most, especially for AI agents.
- Hallucination: the problem thinking does not solve on its own.
- Evals: the way to find out whether a higher effort setting actually improves results on your tasks.
- Claude: the model family whose 3.7 Sonnet release introduced extended thinking by that name.
- Gemini: Google's family of thinking models with thinking budgets.