Recent models can think before they answer. They restate the question, try several approaches, check an intermediate result and drop an approach that does not hold up. This work is done in tokens (chunks of text, often parts of words), just like the answer itself. Reasoning effort sets how much of this work the model does.
One setting, three effects
Raising the effort improves answers on hard tasks: multi-step reasoning, a coding problem, an analysis that has to cross-check information. In return, the answer arrives later and costs more. On a simple task, such as rewording a text or pulling out a date, high effort adds almost nothing.
Key concept
Reasoning effort trades time and money for quality on hard tasks. The right level depends on the task. You measure it on your real cases; you cannot guess it.
Simple task at high effort: 1,500 thinking tokens, about 18 s and 2.25 times the low-effort cost; the answer stays good.
An example from October 2026: at Anthropic, this setting is called effort and has five levels, from low to max. The default is high on most models, and medium on Opus 5.5 and Haiku 5.5. According to the documentation, effort is a signal, not a strict cap: at low effort, the model still thinks about a hard problem, just less. The setting applies to every token the model produces: the thinking, the text of the answer and the tool calls.
An application can also route each request by difficulty: low effort for simple questions, high effort for complex cases. It is the same reasoning as choosing between a small and a large model, which the tokens and cost page details.
What it means for you
For product managers
Not every step of a user journey needs the same effort. Sorting incoming requests has to be instant. A contract analysis can take a minute if the user knows why they are waiting. Decide, step by step, what deserves thinking time, and show that work is under way once the wait runs past a few seconds.
For tech leads
Set the effort explicitly instead of relying on the default, which changes from one model to the next. Compare several levels on your test set every time you change model. In a startup, one level per feature is often enough. In a large company, a platform team can set levels per type of use and track spending per team. If you keep thinking summaries in your logs, protect them like the answers: they can repeat sensitive data from the call.
Go deeperHow much of the thinking you see
In October 2026, on Anthropic's recent models, thinking is on by default and the model decides for itself when to think, and how much, guided by the effort level. By default, the text of that thinking is not returned. When you ask for it, you get a summary of the reasoning, never the raw reasoning. The summary helps you understand an answer. It does not prove the answer is right.
The classic mistake
fAIling RULE I
The trap: Paste your entire codebase into the prompt
The same reflex as pasting everything into the prompt: setting effort to the maximum everywhere, "to be safe," multiplies the bill and the wait on tasks that gain nothing from it. Measure on your real cases, and keep the lowest level that is enough.
Read rule I in the manifestoTry it
Take ten real requests: five simple, five hard. Send them to the same model at low effort, then at high effort. For each one, note the quality of the answer, the wait and the number of tokens billed. Look at where high effort really changes the result.
Around this topic
Go deeperEffort rather than instructions, as of October 2026
According to Anthropic's guide for Claude Opus 5.5, effort is the main control for thinking. To make the model think less, lower the effort first: it works more reliably than a written instruction. In a chat application, lines such as "think carefully before answering" can be removed from the instructions: in Anthropic's tests, replies started sooner with no clear loss of quality. Finally, don't ask the model to copy its internal reasoning into its response: the request may be refused. Ask instead for a short rationale for the answer, or a summary of the actions taken.
Go deeperYou pay for thinking tokens
Invisible does not mean free. At Anthropic, in October 2026, thinking tokens are billed as output tokens, even when their text is not returned. They also count toward the output limit (max_tokens): set it too low and the answer can be cut off before the end. The page Tokens and cost explains what an output token costs.
A blurred thinking strip fills while the answer stays empty; thinking and answer both count as output tokens, and a low limit cuts the answer off. Observed in October 2026 on Anthropic's recent models.
Go deeperWaiting time and streaming
A model that thinks starts writing its answer later. Streaming, which shows the answer word by word as it is written, does not shorten the thinking, but it makes the wait easier to bear once the answer has started. As of October 2026, according to Anthropic's documentation, leaving out the thinking text makes the first word of the answer arrive sooner. That cuts the wait, not the cost.
What to do with it
Decide step by step what deserves thinking time: sort incoming requests instantly, let a contract analysis take a minute.
Set the effort level explicitly instead of relying on the default, which changes from one model to the next.
Compare several levels on your real cases every time you change model, and keep the lowest one that is good enough.