Extended thinking, in plain English.

Anthropic added extended thinking to Sonnet in early 2025. It's a mode where the model generates hidden "thinking" tokens before it writes the visible reply. Think of it as the model working through the problem on scratch paper before drafting the answer.

LADLE exposes this as a toggle in the composer. The default is off. Here's the honest guide to using it.

**What extended thinking does.** When it's on, Sonnet allocates a budget of tokens (LADLE budgets 8,000) to internal reasoning before generating the reply. The reasoning is captured and shown in a collapsed "Thinking…" block above the reply — you can expand it to see how the model got there. The reply itself is often better on hard problems: fewer arithmetic mistakes, tighter logic, better handling of multi-step questions.

**What it doesn't do.** It doesn't make the model smarter. It gives an already-capable model more room to check its own work. On easy questions, the extra thinking is wasted (and costs you tokens against your budget). On medium-hard questions, the reply tends to be a hair more careful, sometimes noticeably so. On genuinely hard questions — multi-step math, complex logic puzzles, chained reasoning problems — it makes a real difference.

**When to turn it on:** - Math or logic problems where the answer is verifiable and the intermediate steps matter - Multi-step reasoning where the model needs to keep track of state across the argument - Complex code where getting the algorithm right in one shot matters more than iterating - Any task where you'd naturally say "take your time to get this right"

**When to leave it off:** - Drafting, editing, translation — these don't need extra reasoning, they need voice - Short factual questions — the reply is done before extended thinking would even finish - Conversational back-and-forth — the extra latency (thinking tokens generate before the visible reply starts) makes the chat feel slow - Anything where the answer is stylistic rather than logical

**On latency.** Extended thinking makes replies take longer. Not a lot longer — usually 5-15 seconds extra on a typical turn — but noticeable. If you're in a flow of quick turns, that friction adds up. Turn it on for the hard one; leave it off for the flow.

**On cost.** Thinking tokens are billed at the same rate as output tokens (Sonnet: $15/million). An 8,000-token thinking budget adds about $0.12 of compute per turn. Not huge on its own but real if you leave it on across many turns. LADLE's metering counts thinking against your monthly budget the same as output tokens.

**On seeing the reasoning.** The "Thinking…" block is expandable. Some users find it fascinating; others find it noise. It's collapsed by default. What's worth knowing: the reasoning shown is a compressed representation of the model's process, not a raw transcript — Anthropic surfaces enough for it to be useful without cluttering the reply.

**Rule of thumb.** If you'd naturally say "take your time," turn it on. If you want the reply now, leave it off. The toggle is per chat, so a heavy-reasoning research chat can have it on while your daily drafting chat has it off.