Thinking & effort

Modern Claude models can think before they answer: reasoning step by step internally before producing a response. Understanding this helps you get better answers on hard problems.

What "thinking" means

When faced with a complex task, Claude can do extended (or adaptive) thinking: it works through the problem privately, considers approaches, checks itself, and only then writes the final answer. This dramatically improves accuracy on:

• Multi-step math and logic.
• Complex planning and trade-off analysis.
• Tricky code and debugging.
• Careful reading of long, dense material.

On recent models this is often adaptive: Claude decides how much to think based on difficulty, so you don't have to micromanage it.

When it helps (and when it doesn't)

Good fit for more thinking	Doesn't need it
"Find the flaw in this argument."	"Rewrite this sentence."
"Plan a migration with dependencies."	"Give me 5 synonyms."
"Reconcile these conflicting numbers."	"Format this as a table."

For simple tasks, heavy thinking just adds latency and cost without improving the result.

The effort lever (for builders)

On the API, recent Opus and Sonnet models expose an effort parameter that trades intelligence for latency and cost within a single model. Turning effort up can be a better move than switching to a bigger model.

low effort   → faster, cheaper, fine for routine tasks
high effort  → slower, pricier, better on hard reasoning

TIP

Tuning effort is often a smarter lever than jumping to a more expensive model. Try high effort on Sonnet before reaching for Opus.

How to invite better thinking in everyday chat

Even without API parameters, your prompt can encourage deeper reasoning:

Think this through carefully before answering.
First, list the constraints. Then evaluate each option
against them. Only then give your recommendation.

Asking Claude to show its reasoning and consider alternatives often yields a more reliable answer, especially for decisions and analysis.

Don't over-prompt simple tasks

For trivial requests, telling Claude to "think step by step" just makes it slower and more verbose. Match the effort to the difficulty.

You now understand the models. The single biggest skill is next: prompting.