Foundations
Extended thinking
For hard problems, Claude can do its reasoning out loud in a dedicated
thinking phase before it answers — a scratchpad you can budget with tokens. More thinking buys
accuracy on complex tasks, but costs tokens and latency. Slide the thinking budget and see the trade-off.
thinking budgetreasoning
accuracy vs costwhen to enable
Explain like I'm 10
On a hard maths problem, you do better if you
work it out on scratch paper instead of blurting the
first number. Extended thinking gives Claude scratch paper. For easy questions ("what's the capital of
France?") it's a waste of time — just answer. For tricky multi-step puzzles, the scratch work is what gets it
right. You decide how much paper to hand over.
Dial the thinking budget
Below, the same hard planning task is solved with different thinking budgets. Watch accuracy climb — then
flatten — while cost and latency keep rising. The art is finding "enough," not "maximum":
When to turn it on (and off)
| Enable thinking for… | Skip it for… |
|---|
| multi-step math, logic, planning | simple lookups & classification |
| complex agentic decisions / tool orchestration | high-volume, latency-sensitive endpoints |
| hard debugging or architecture trade-offs | tasks a smaller model already nails |
| anything where a wrong answer is expensive | creative/format tasks that don't need reasoning |
Exam trap: extended thinking isn't free and isn't always better — for easy or
high-throughput tasks it just adds cost and latency. Match it to problem difficulty, and size the budget
with an eval rather than maxing it out. It also pairs with model choice: a hard task might be Sonnet +
thinking, or Opus — decide by measuring, not by reflex.
Takeaways: extended thinking gives Claude a token-budgeted scratchpad
before answering — real accuracy gains on hard, multi-step problems, wasted cost/latency on easy ones.
Returns flatten past "enough," so tune the budget with evals and reserve it for difficult or high-stakes tasks.
Curated companion: Anthropic — Extended thinking.