NastępnyNastępny poradnik
Jak tworzyć prompty dla modeli rozumujących
Językowa sztuczna inteligencja
PRZEWODNIK Językowy AI
Wysiłek rozumowania i budżety myślowe to ustawienia modeli rozumowania, które kontrolują ilość wewnętrznej pracy krok po kroku wykonywanej przez model przed udzieleniem odpowiedzi.
Posłuchaj tego przewodnikaOpowiadane przez sztuczną inteligencję · ~3 min
More thinking often improves results on hard problems like math, coding and multi-step analysis, but it adds cost and delay, and past a point it stops helping. Choosing the right level per task is one of the simplest ways to balance quality, speed and price.
Reasoning models are trained to produce a chain of intermediate reasoning before their final answer. That reasoning is made of tokens, just like the answer, and generating it takes time and money. Providers let you control how much of it happens. OpenAI exposes a reasoning_effort parameter on its reasoning models with levels such as low, medium and high; some newer models add a minimal level. Anthropic's extended thinking on supported Claude models takes a budget_tokens value that sets the maximum number of tokens Claude may use for thinking; the budget has a minimum of 1,024 tokens and must be smaller than max_tokens. Google's Gemini 2.5 models accept a thinking budget, and some can have thinking turned down or off. The names differ, but the idea is the same: a dial between fast and thorough. These settings are guidance, not exact counts. A model given a large budget may use only part of it on an easy question, and effort levels shift behavior rather than fixing a number. Cost is the part people often miss. Reasoning tokens are generally billed as output tokens even when you cannot see them, and output tokens usually cost more than input tokens. A short visible answer can carry a large hidden bill. Latency grows too, since every reasoning token is generated in sequence. More thinking has diminishing returns. On easy or factual lookup questions, extra reasoning rarely improves accuracy and can sometimes lead a model to overcomplicate or second-guess a correct answer. Gains are largest on problems that genuinely need several steps. A common misconception is that the highest setting is always best. The practical approach is to test your own tasks at several levels and pick the lowest one that meets your quality bar.
A customer support bot answering simple order questions uses low reasoning effort, which keeps replies fast and cheap without hurting accuracy.
A developer debugging a tricky concurrency bug switches to high effort for that one request and gets a correct diagnosis the low setting missed.
A team using Anthropic's extended thinking sets a budget of a few thousand tokens for routine document review and a much larger budget only for contract clauses flagged as unusual.
An analyst runs the same set of 50 test questions at low, medium and high effort and finds medium matches high on accuracy for their tasks at a lower cost, so they standardize on medium.
Providers are moving toward models that decide for themselves how long to think, with effort settings acting as bounds rather than fixed instructions; Anthropic's newer models, for example, offer adaptive thinking, and some products already switch automatically between fast and reasoning modes. Research continues into making reasoning more efficient so that similar accuracy needs fewer tokens. Parameter names, allowed values and pricing have changed several times and will likely keep changing, so treat current settings as provider-specific details and re-run your own evaluations when you upgrade models.
Darmowy newsletter
Trzy zweryfikowane historie AI każdego poranka w dni robocze, napisane prostym angielskim. Bezpłatne na zawsze, bez reklam.
Jedno e‑mail każdego dnia roboczego. Wypisz się jednym kliknięciem. Nigdy nie sprzedajemy ani nie udostępniamy Twojego adresu.
Sprawdź się
Natychmiastowa informacja zwrotna na każde pytanie oraz udostępnialny certyfikat z weryfikowalnym ID po zaliczeniu kursu.
Wspieraj darmową edukację AI. AI Understanding jest organizacją non-profit 501(c)(3) — bez reklam, bez płatnej ściany, zawsze. Przekaż darowiznę
Reasoning effort and thinking budgets are settings on reasoning models that control how much internal step-by-step work the model does before answering. More thinking often improves results on hard problems like math, coding and multi-step analysis, but it adds cost and delay, and past a point it stops helping. Choosing the right level per task is one of the simplest ways to balance quality, speed and price.
These settings adjust how much intermediate reasoning a reasoning model generates before its final answer.
Reasoning tokens are generally billed as output tokens, which usually cost more than input, even when they are not shown.
The budget has a minimum of 1,024 tokens and must be below max_tokens so there is room for the answer.
Budgets are upper limits and guidance. On easy questions the model often uses much less.
Gains are largest on multi-step problems. Easy tasks rarely benefit and can even suffer from overthinking.
Ucz się dalej
Wybrano więcej przewodników na ten temat
NastępnyNastępny poradnik
Jak tworzyć prompty dla modeli rozumujących
Językowa sztuczna inteligencja