Extended thinking (THINK).

Extended thinking is Anthropic's mode where the model uses a private reasoning scratchpad before emitting its visible answer. LADLE exposes it as a per-chat toggle. Turn it on with the THINK chip in the composer's tools popover; it persists on the chat until you turn it off.

**Sonnet-only:** Extended thinking is a Sonnet feature; Haiku doesn't support it. Turning THINK on forces the chat's routing to Sonnet (heavy tier) for the rest of the conversation. The reason chip on each assistant reply will read "user selected THINK → heavy."

**Budget:** 8,000 tokens are allocated for the reasoning trace on every turn. Typical actual usage is 1,500-4,500 tokens for a substantive turn. Short questions use less.

**Cost:** Thinking tokens are billed as OUTPUT tokens at Sonnet's rate ($15 per million). A typical THINK-enabled turn costs roughly twice what the same turn would cost with THINK off, for the same visible reply length.

**How it renders:** Before the visible reply starts, you'll see a collapsed "THINKING… N chars" chip above the reply. When the reply finishes, expand the chip to inspect the full reasoning trace. The trace is stored persistently, so it's still there when you re-open the chat later.

**When it's worth turning on:** - Math and logic problems - Multi-step reasoning tasks - Debugging that requires ruling out several hypotheses in order - Careful reading of a long document with specific criteria

**When it's overkill:** - Creative writing (the reasoning trace becomes stilted) - Casual conversation - Simple lookups

The doubled latency and cost pay off on hard reasoning problems and are wasted on easy ones. Toggle it per chat, not globally.