What Is a Reasoning Model? And When to Turn Thinking Off
Your assistant has a switch on it now. ChatGPT calls it Thinking, Claude calls it extended thinking, Gemini calls it Extended Thinking. In all three the fast mode is what you get unless you ask for the other one.
The obvious assumption is that the thinking one is the better one, and you leave it on when the work matters. That is half right, and the half that is wrong costs you accuracy rather than just time.
What the model is doing when it “thinks”
Nothing mysterious, and nothing like deliberation.
A reasoning model writes before it answers you. It generates a long run of intermediate text working through the problem, then produces the response you see. That intermediate text is a The model’s step-by-step reasoning, written out as ordinary text before the final answer. It matters because each step becomes part of what the model is reading when it writes the next one, so putting a result down on the page makes it available to everything that follows. It is closer to showing your work on paper than to thinking quietly and then speaking. , and the reason it helps at all is mechanical rather than mystical. The model is always predicting what comes next from what is already on the page. Put a correct intermediate step on the page and the next step has something true to build on.
What it is not is a switch marked “try harder.” The model is not holding back effort in fast mode. It is producing a different quantity of text before committing to an answer, and more text is not reliably better text.
Three levels of difficulty, and only one wants thinking
The most useful finding here comes from a team at Apple, who tested reasoning models against ordinary ones across a range of difficulty. They found three distinct bands:
(1) low-complexity tasks where standard models outperform LRMs, (2) medium-complexity tasks where LRMs demonstrates advantage, and (3) high-complexity tasks where both models face complete collapse.
On easy tasks the ordinary model was more accurate than the reasoning model, not merely faster. It contradicts the intuition that thinking is a better setting you pay for in seconds.
They also found something stranger at the top end. As problems got harder the models produced longer chains of thought, up to a point, and then produced shorter ones even though they had plenty of room left. Effort went up and then quietly went down again, right where you would most want it.
Overthinking is measured, not anecdotal
A separate group built a tool to solve exactly this. Their opening line is the finding: reasoning models “excel at solving complex tasks,” but this “often leads to overthinking on simple tasks, resulting in unnecessary computational overhead.”
Their system watches the task and switches between short and long chains automatically, and it cuts computation by 20 to 30% while holding accuracy on the genuinely hard problems. The result is worth less to you than what it implies: engineers are building machinery whose entire job is to turn thinking off when it is not earning its keep. Nobody builds that for a setting that is free to leave on.
The collapse finding is contested, and that matters
One caution before you take all of that as settled.
The part everyone quoted was the third band, where both kinds of model fail completely. Other researchers pushed back hard on it. A commentary by A. Lawsen argues the collapse “primarily reflect[s] experimental design limitations rather than fundamental reasoning failures,” on three grounds: the Tower of Hanoi tests could exceed the models’ output token limits, with models saying so in their own answers; the automated scoring could not tell a reasoning failure apart from a practical constraint; and some river-crossing puzzles were “mathematically impossible instances” that models were marked wrong for failing to solve.
Those are specific objections, not vague complaints. So treat the collapse on hard problems as an open question rather than a fact.
They do much less damage to the first two bands, and the reason is worth following. Those two are comparisons: the same tasks, marked the same way, with the ordinary model doing better on the easy end and the reasoning model doing better in the middle. If the marking was unfair, it was unfair to both models equally, so it cannot be the reason one of them beat the other.
There is a second caveat worth knowing, which is what the tests were made of. The researchers used Artificial problems like Tower of Hanoi, where the difficulty can be dialed up one notch at a time while the logic stays identical. That control is the whole point: it lets you find the exact rung where performance changes, which a pile of mixed real-world tasks never could. rather than real work, and they did it deliberately, because the usual math and coding benchmarks suffer from When the test questions, or near-copies of them, were in the material the model was trained on. The model then scores well by having effectively seen the answer, so the benchmark measures memory rather than capability. It is the central difficulty in evaluating these systems, and it is why a fresh puzzle nobody has published tells you more than a famous exam does. . The cost of that choice is that a puzzle is not a quarterly report. Take the general pattern from this, which is that easy work suits the plain model and middling work suits the reasoning one. Do not try to work out where your own tasks sit on their difficulty scale, because nobody has run this experiment on the work in your inbox.
So when do you switch it on
The practical rule follows from what a chain of thought does. It helps when there are intermediate results that later steps depend on, and it has nothing to work with when there are not.
Turn thinking on when the task has steps that can be individually wrong:
- Arithmetic or anything with a calculation inside it, especially several chained together.
- Comparing options against more than two or three criteria at once.
- Planning where the order matters, or where one choice constrains another.
- Anything where you would ask a colleague to show their work.
Leave it off for tasks that are one thing done well:
- Rewriting, shortening, changing tone, fixing grammar.
- Summarizing something short.
- Formatting, reformatting, converting a list into a table.
- Pulling a specific detail out of a document you supplied.
A quick test: does the task have middle steps? If getting the final answer right depends on getting two or three smaller things right first, turn thinking on. If the task is really one judgment about wording, extra steps only give the model more chances to talk itself into something else.
Do check the setting rather than assuming. In all three products the fast mode is the default, and the names and defaults move around between releases.
What thinking does not fix
This is the part worth remembering, because it is where people are most likely to be misled by a good idea.
Thinking mode does not fix invented facts. A model makes things up because certain facts were never learnable and because guessing scores better than abstaining in their benchmark testing, and neither of those changes when you let it write more first. What you get instead is a made-up answer with step-by-step reasoning laid out underneath it. That makes it more convincing. It does not make it true.
It also does not fix missing context. If the model does not know your constraint, more reasoning over the same gap produces a longer, better argued answer to the wrong question. Reasoning is applied to what is in front of it, so what you put in front of it is still the thing that decides the answer.
The takeaway
Thinking mode is not a quality dial. It is a match between one kind of task and one kind of processing.
Switch it on when the answer has steps, and off when the answer is one judgment. On easy work the plain model tested more accurate, not just quicker, and a longer chain of reasoning over facts the model invented or never had is not a better answer. It is a more persuasive one.
Sources: the three difficulty bands and the declining-effort finding are from Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh, Maxwell Horton, Samy Bengio and Mehrdad Farajtabar of Apple, The Illusion of Thinking (opens in a new tab) (arXiv 2506.06941, 7 June 2025, revised November 2025). The rebuttal is A. Lawsen’s Comment on The Illusion of Thinking (opens in a new tab) (arXiv 2506.09250, June 2025). The overthinking finding and the 20 to 30% figure are from Guosheng Liang, Longguang Zhong, Ziyi Yang and Xiaojun Quan, ThinkSwitcher (opens in a new tab) (arXiv 2505.14183, 20 May 2025). All three are read from their abstracts: the full PDFs would not extract, so I have not checked the experiments themselves, and none of the three is peer reviewed. Both the Apple paper and ThinkSwitcher test specific model versions from 2025, and this field moves. The routing rule, the single-judgment test, and the argument that thinking mode does not fix hallucination or missing context are mine, reasoned from what a chain of thought does rather than measured by any of these papers.
Related: Why AI Makes Things Up: It Was Trained to Guess is the failure that more reasoning cannot repair. Should You Hand AI the Whole Task? The People Who Know It Best Don’t is the same question one level up, about how much of the job to hand over rather than which mode to hand it to.