Extended thinking: when to enable.

LADLE's Deep composer has a thinking toggle. When on, Sonnet generates hidden reasoning tokens before writing the visible reply. The reasoning shows in a collapsed "Thinking…" block above the answer — expand it to see how the model got there.

What it does

Extended thinking gives the model scratch paper. On hard multi-step problems (complex math, logic puzzles, multi-hop reasoning, code that has to be correct in one shot), the extra reasoning meaningfully improves accuracy. Fewer arithmetic errors, tighter argumentation, better handling of tricky edge cases.

The mechanics: Anthropic reserves a thinking-token budget (LADLE sets 8,000) that the model uses before writing the visible response. The visible answer follows.

When to enable

- **Math or logic problems** where the intermediate steps matter and the answer is verifiable. - **Multi-step reasoning** where the model has to keep track of state across the argument. - **Complex code** where getting the algorithm right in one shot beats iterating. - **Any task where you'd naturally say "take your time."**

When to leave off

- **Drafting, editing, translation.** These don't need extra reasoning — they need voice. Thinking can actually make the reply feel more constructed and less natural. - **Short factual questions.** The reply is done before extended thinking would even finish. - **Conversational back-and-forth.** Extra latency (thinking tokens generate first, then visible reply) makes chat feel slow.

Latency cost

Extended thinking adds 5-15 seconds to typical turns. Not huge, but noticeable in a flow. Turn on for the hard turn; leave off for the quick ones.

Token cost

Thinking tokens bill at Sonnet's output rate ($15/MTok). An 8,000-token budget adds ~$0.12/turn max. LADLE meters thinking against your monthly budget as output tokens.

What Anthropic ships vs what LADLE surfaces

Anthropic ships extended thinking as a per-request parameter. LADLE surfaces it as a per-chat composer toggle — flip it on for a research or coding chat, leave it off for daily use. The setting persists per chat, so a research chat stays "thinking-on" without you re-enabling every turn.

The reasoning block

The Thinking… block above each reply is collapsed by default. Some users find it fascinating (watching the model's actual chain of thought). Others find it noise. Either is fine.

Note: the reasoning shown is a compressed representation of the model's process, not a raw dump — Anthropic surfaces enough to be useful without cluttering the reply.