Lesson 5 · 7 min
Extended Thinking
When to enable thinking budgets and what they cost.
Extended thinking gives Claude scratchpad tokens before producing the final response. Use it for hard reasoning (math, multi-step planning, code analysis). Bill is on thinking tokens too; budget explicitly.
Production scenario
Real-world example: Math tutoring on word problems
A K-12 math tutor handles two query types: arithmetic lookups ("what's 234 × 7?") and word problems ("Lisa has three times as many marbles as Tom..."). The first needs no thinking budget. The second does.
const thinking = isWordProblem(question) ? { budget_tokens: 4000 } : undefined;
const reply = await claude.messages.create({
model: "claude-opus-4-7",
thinking,
system: TUTOR_PROMPT,
messages: [{ role: "user", content: question }],
});Pass rate on word problems jumps from 71% to 89% with extended thinking. Lookups stay fast and cheap.
Why this matters: extended thinking is worth its tokens on genuinely multi-step reasoning. Don't sprinkle it on everything — gate it on the question shape.
Knowledge points in this lesson
- Extended thinking gives a scratchpad budget
- Bill is paid in thinking tokens too
- Best for hard multi-step reasoning
- Overkill on simple lookup tasks
- Budget thinking explicitly per call
Quick check
Prompt EngineeringSelect one
How many few-shot examples typically peak performance?
