{"course":{"slug":"prompting","title":"Getting Real Work out of a Model","url":"https://addaly.com/learn/prompting","level":"beginner","track":"start"},"lesson":{"slug":"reasoning-models","title":"Models that think before answering","url":"https://addaly.com/learn/prompting/reasoning-models","position":40,"of":72,"minutes":9,"answer":"Reasoning models pay off exactly where you would have asked for steps: state the goal and constraints, not the method.","keyIdea":"Reasoning models pay off exactly where you would have asked for steps: state the goal and constraints, not the method, and treat the visible trace as material to check rather than a record of what happened."},"body":{"format":"markdown","text":"## What they actually do\n\nA reasoning model generates a long internal working-out before producing its answer. You are usually shown a summary of that working, or nothing at all, and you pay for the hidden tokens. Some products expose a control (effort, thinking budget, extended thinking) that trades money and latency for more internal work.\n\nThis is not a different architecture doing something mysterious. It is the same next-token machinery, trained so that producing a long chain of intermediate reasoning before answering improves the answer, and given room to do it.\n\n## Where they are worth it\n\nGenuinely better on multi-step problems: mathematics, logic puzzles, competitive programming, debugging, planning under several constraints, and picking apart an argument. On these the difference is not marginal.\n\nNot better, and often worse value, on: recall, summarising, writing, tone work, translation, or anything where the answer is retrieved rather than derived. You pay several times the cost and wait several times as long for a paragraph you could have had immediately, sometimes with a stiffer register.\n\nThe practical rule: **if the task would have benefited from you asking for steps, use a reasoning model. If it would not, do not.**\n\n## How prompting changes\n\nThree things are different, and the first surprises people.\n\n**Stop telling it to think step by step.** It already is. The instruction is redundant, and providers of these models say so explicitly. In some cases an elaborate reasoning scaffold in the prompt interferes with the model's own process and makes things worse.\n\n**Give the goal and the constraints, not the method.** With an ordinary model you often specify the route: first do this, then that. With a reasoning model, state the destination clearly, state what a good answer must satisfy, and let it find the route. Over-specifying the method is the most common way people get less out of these models than they should.\n\n**Be complete about the requirements.** Reasoning models exploit constraints. A well-stated constraint gives the internal process something to check against; a missing one is a degree of freedom that produces a confident wrong answer more elaborately than before.\n\n## The honest part about the visible reasoning\n\nYou will often see a summary of the thinking. It is tempting to read it as a log of what happened. It is not.\n\nThe displayed trace is a summary, sometimes produced by a separate model, and providers have been fairly open that it is not a faithful transcript of the internal process. More uncomfortably, research on chain-of-thought faithfulness (including work published by model developers themselves) has found that models often fail to mention information they demonstrably used, such as a hint planted in the prompt that changed the answer.\n\nSo the same rule as before holds, with more force: shown reasoning is material to check, not proof of how the answer was reached. A trace that looks careful is not evidence that the answer is right, and a trace containing an error under a correct answer means both need looking at.\n\n## Cost and latency, concretely\n\nA reasoning model on a hard problem can generate thousands of hidden tokens before its first visible word. That means:\n\n- **You wait.** Ten seconds to several minutes, depending on the model and the effort setting.\n- **You pay for tokens you never see.** On metered APIs this dominates the bill.\n- **Batch jobs multiply it.** A reasoning model over 500 documents is a different budget from a fast model over 500 documents.\n\nFor a personal chat this is irrelevant. For anything running repeatedly, it decides your architecture, and the sensible pattern is a fast model for the routine cases and a reasoning model for the ones flagged as hard.\n\n## Overthinking is real\n\nOn easy questions, more thinking is not better and sometimes measurably worse: the model reconsiders a correct answer into a wrong one. If your effort setting is adjustable, low or medium handles most work. Reserve high for problems you could not solve quickly yourself.\n\n## What to do this week\n\nTake three tasks you do often. Run each on a fast model and a reasoning model, twice each, and look at the difference against the time you waited.\n\nYou will usually find one task where the reasoning model is transformative and two where it is a slower way to get the same paragraph. That is the entire decision, and it is specific to your work rather than to anybody's benchmark."},"check":{"question":"An engineer moves a document-summarising job to a reasoning model at high effort. Quality is unchanged, the bill triples, and each response takes 40 seconds. What is the correct reading?","options":[{"text":"The prompt needs restructuring to take advantage of the model's reasoning.","correct":false,"tempting":"Assumes the capability fits the task; summarising is retrieval and compression, not derivation."},{"text":"Quality gains from reasoning models appear only over long time horizons.","correct":false,"tempting":"Not a real effect; the gains are per-task and immediate where they exist at all."},{"text":"The effort setting should be raised further to see a benefit.","correct":false,"tempting":"Increases the cost of the thing already shown not to help, and overthinking can degrade easy tasks."},{"text":"Summarising has no dependent intermediate results to derive, so the extra internal work bought nothing and the fast model was the right tool.","correct":true}]},"prev":{"slug":"arithmetic-and-code","title":"Arithmetic it should not do in its head","url":"https://addaly.com/learn/prompting/arithmetic-and-code"},"next":{"slug":"breaking-a-task-into-steps","title":"Splitting a task that keeps failing","url":"https://addaly.com/learn/prompting/breaking-a-task-into-steps"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Models that think before answering', in Getting Real Work out of a Model. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/prompting/reasoning-models (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Models that think before answering","courseTitle":"Getting Real Work out of a Model","module":{"index":5,"slug":"making-it-work-the-problem","title":"Making it work the problem"},"answer":null,"keyIdea":"Reasoning models pay off exactly where you would have asked for steps: state the goal and constraints, not the method, and treat the visible trace as material to check rather than a record of what happened.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"what-they-actually-do","runs":[{"text":"What they actually do"}]},{"t":"p","runs":[{"text":"A reasoning model generates a long internal working-out before producing its answer. You are usually shown a summary of that working, or nothing at all, and you pay for the hidden tokens. Some products expose a control (effort, thinking budget, extended thinking) that trades money and latency for more internal work."}]},{"t":"p","runs":[{"text":"This is not a different architecture doing something mysterious. It is the same next-token machinery, trained so that producing a long chain of intermediate reasoning before answering improves the answer, and given room to do it."}]},{"t":"h2","id":"where-they-are-worth-it","runs":[{"text":"Where they are worth it"}]},{"t":"p","runs":[{"text":"Genuinely better on multi-step problems: mathematics, logic puzzles, competitive programming, debugging, planning under several constraints, and picking apart an argument. On these the difference is not marginal."}]},{"t":"p","runs":[{"text":"Not better, and often worse value, on: recall, summarising, writing, tone work, translation, or anything where the answer is retrieved rather than derived. You pay several times the cost and wait several times as long for a paragraph you could have had immediately, sometimes with a stiffer register."}]},{"t":"p","runs":[{"text":"The practical rule: "},{"text":"if the task would have benefited from you asking for steps, use a reasoning model. If it would not, do not.","b":true}]},{"t":"h2","id":"how-prompting-changes","runs":[{"text":"How prompting changes"}]},{"t":"p","runs":[{"text":"Three things are different, and the first surprises people."}]},{"t":"p","runs":[{"text":"Stop telling it to think step by step.","b":true},{"text":" It already is. The instruction is redundant, and providers of these models say so explicitly. In some cases an elaborate reasoning scaffold in the prompt interferes with the model’s own process and makes things worse."}]},{"t":"p","runs":[{"text":"Give the goal and the constraints, not the method.","b":true},{"text":" With an ordinary model you often specify the route: first do this, then that. With a reasoning model, state the destination clearly, state what a good answer must satisfy, and let it find the route. Over-specifying the method is the most common way people get less out of these models than they should."}]},{"t":"p","runs":[{"text":"Be complete about the requirements.","b":true},{"text":" Reasoning models exploit constraints. A well-stated constraint gives the internal process something to check against; a missing one is a degree of freedom that produces a confident wrong answer more elaborately than before."}]},{"t":"h2","id":"the-honest-part-about-the-visible-reasoning","runs":[{"text":"The honest part about the visible reasoning"}]},{"t":"p","runs":[{"text":"You will often see a summary of the thinking. It is tempting to read it as a log of what happened. It is not."}]},{"t":"p","runs":[{"text":"The displayed trace is a summary, sometimes produced by a separate model, and providers have been fairly open that it is not a faithful transcript of the internal process. More uncomfortably, research on chain-of-thought faithfulness (including work published by model developers themselves) has found that models often fail to mention information they demonstrably used, such as a hint planted in the prompt that changed the answer."}]},{"t":"p","runs":[{"text":"So the same rule as before holds, with more force: shown reasoning is material to check, not proof of how the answer was reached. A trace that looks careful is not evidence that the answer is right, and a trace containing an error under a correct answer means both need looking at."}]},{"t":"h2","id":"cost-and-latency-concretely","runs":[{"text":"Cost and latency, concretely"}]},{"t":"p","runs":[{"text":"A reasoning model on a hard problem can generate thousands of hidden tokens before its first visible word. That means:"}]},{"t":"ul","items":[[{"text":"You wait.","b":true},{"text":" Ten seconds to several minutes, depending on the model and the effort setting."}],[{"text":"You pay for tokens you never see.","b":true},{"text":" On metered APIs this dominates the bill."}],[{"text":"Batch jobs multiply it.","b":true},{"text":" A reasoning model over 500 documents is a different budget from a fast model over 500 documents."}]]},{"t":"p","runs":[{"text":"For a personal chat this is irrelevant. For anything running repeatedly, it decides your architecture, and the sensible pattern is a fast model for the routine cases and a reasoning model for the ones flagged as hard."}]},{"t":"h2","id":"overthinking-is-real","runs":[{"text":"Overthinking is real"}]},{"t":"p","runs":[{"text":"On easy questions, more thinking is not better and sometimes measurably worse: the model reconsiders a correct answer into a wrong one. If your effort setting is adjustable, low or medium handles most work. Reserve high for problems you could not solve quickly yourself."}]},{"t":"h2","id":"what-to-do-this-week","runs":[{"text":"What to do this week"}]},{"t":"p","runs":[{"text":"Take three tasks you do often. Run each on a fast model and a reasoning model, twice each, and look at the difference against the time you waited."}]},{"t":"p","runs":[{"text":"You will usually find one task where the reasoning model is transformative and two where it is a slower way to get the same paragraph. That is the entire decision, and it is specific to your work rather than to anybody’s benchmark."}]}],"check":{"question":[{"text":"An engineer moves a document-summarising job to a reasoning model at high effort. Quality is unchanged, the bill triples, and each response takes 40 seconds. What is the correct reading?"}],"options":[{"text":[{"text":"The prompt needs restructuring to take advantage of the model’s reasoning."}],"correct":false,"tempting":[{"text":"Assumes the capability fits the task; summarising is retrieval and compression, not derivation."}]},{"text":[{"text":"Quality gains from reasoning models appear only over long time horizons."}],"correct":false,"tempting":[{"text":"Not a real effect; the gains are per-task and immediate where they exist at all."}]},{"text":[{"text":"The effort setting should be raised further to see a benefit."}],"correct":false,"tempting":[{"text":"Increases the cost of the thing already shown not to help, and overthinking can degrade easy tasks."}]},{"text":[{"text":"Summarising has no dependent intermediate results to derive, so the extra internal work bought nothing and the fast model was the right tool."}],"correct":true,"tempting":null}],"order":[2,1,3,0],"why":null},"prev":{"slug":"arithmetic-and-code","title":"Arithmetic it should not do in its head","minutes":9},"next":{"slug":"breaking-a-task-into-steps","title":"Splitting a task that keeps failing","minutes":9}}}