# Extended thinking

An optional mode where Claude produces internal reasoning before its answer, improving hard tasks at extra token cost.

## Why the CCAR-P exam tests this

The exam checks you know extended thinking helps complex reasoning but costs tokens and latency, so it is not free to enable everywhere.

## Definition

Extended thinking lets Claude spend additional **reasoning tokens** working through a problem before it commits to a final answer. On genuinely hard tasks — multi-step math, complex planning, tricky debugging — this measurably improves quality by letting the model decompose and check its work.

The cost is real: those thinking tokens are **billed** and add **latency**, and you often control a **thinking budget** that caps how much reasoning is allowed. Enabling it for simple lookups or short classifications wastes both. The right pattern is to turn it on for tasks whose difficulty justifies the extra deliberation, and leave it off for fast, simple, high-volume calls.

Extended thinking is a per-request capability, orthogonal to the model tier — you choose whether to enable it and how large a budget to allow. More thinking is not automatically better; past a point it adds cost without improving the answer.

## Exam trap

Extended thinking is not free — it consumes billed tokens and adds latency, so "always enable thinking" is wrong for simple or latency-sensitive tasks.

## Commonly confused with

- **model-selection-opus-sonnet-haiku** — Thinking is a mode you toggle per request, separate from which tier you pick.
- **prompt-caching** — Caching saves input reprocessing; thinking adds output/reasoning tokens — opposite effects on token count.

## Related

- model-selection-opus-sonnet-haiku
- context-window-management
- structured-output

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P2). Not affiliated with Anthropic.