Claude Extended Thinking

Extended thinking gives Claude dedicated thinking time before it produces its final answer. Instead of responding immediately, Claude works through the problem step by step in a thinking block - exploring options, checking assumptions, considering edge cases - and only then writes its response. This is Claude at its most powerful: the same model that handles routine questions instantly can also tackle genuinely complex reasoning tasks that require holding many variables in mind simultaneously. Think of it as the difference between an analyst who blurts out the first thing that comes to mind versus one who works through the numbers carefully before presenting a recommendation.

You enable extended thinking by passing thinking={"type": "enabled", "budget_tokens": N} to the messages call, where budget_tokens is the maximum number of tokens Claude can use for its internal reasoning (not billed as output - only the final response is billed as output tokens). The response contains a thinking block (Claude's reasoning, which you can show or hide in your UI) followed by a text block (the final answer). Higher budget_tokens enables deeper reasoning but increases latency and cost. For most complex business tasks, a budget of 5,000-10,000 tokens strikes the right balance.

The below example shows a wealth management firm using extended thinking for a complex investment allocation decision - comparing risk, returns, liquidity, and tax implications across multiple asset classes before making a recommendation.

It gives the following output showing thinking then the final recommendation,

=== CLAUDE'S REASONING PROCESS (Extended Thinking) ===
Let me work through this carefully. Priya is 38, wants to retire at 55 - that's
17 years. She needs Rs 2L/month in retirement. Assuming 6% inflation and 25 years
of retirement, I need to calculate the corpus needed...

Retirement corpus needed: Rs 2L x 12 x 25 years + inflation adjustment...
At 6% inflation over 17 years, Rs 2L today = ~Rs 5.4L at retirement...
Using 4% safe withdrawal rate: corpus needed = Rs 5.4L x 12 / 0.04 = Rs 16.2 crores

Current position: Rs 80L savings + Rs 30L EPF = Rs 1.1 crore
Shortfall: Rs 16.2 crore - Rs 1.1 crore = Rs 15.1 crore to build in 17 years...

Her Rs 1.5L/month surplus: she needs aggressive but not reckless growth...
[reasoning continues for 8,000 tokens...]

=== INVESTMENT RECOMMENDATION ===
Priya, here is your personalised 17-year wealth plan:

IMMEDIATE (Next 3 months):
- Move Rs 5L to liquid fund for emergency corpus (Parag Parikh Liquid Fund)
- Start SIP: Rs 75,000/month in Nifty 50 index fund (UTI or Mirae Asset)
- Start SIP: Rs 40,000/month in flexi-cap fund (Parag Parikh Flexi Cap)
- Start SIP: Rs 20,000/month in international equity fund (Motilal S&P 500)
- PPF contribution: Rs 12,500/month (Rs 1.5L annually for 80C benefit)
- Remaining Rs 2,500: top up NPS Tier 1 (additional Rs 50,000 tax benefit u/s 80CCD)

EDUCATION GOAL (8 years, Rs 25L target):
- Start dedicated SIP of Rs 15,000/month in Sukanya Samriddhi (if daughter <10)
  or balanced advantage fund with 8-year horizon
- Projected corpus at 12% CAGR: Rs 26.8L - goal achieved

PROJECTED RETIREMENT CORPUS at 55: Rs 15.8 crores at blended 12% CAGR
This covers your Rs 2L/month retirement income with a comfortable 20% buffer.

Extended thinking is worth the extra cost when the answer genuinely requires multi-step reasoning that builds on itself: financial planning across multiple goals and time horizons, legal analysis weighing competing clauses, complex pricing decisions with many interdependent variables, and medical differential diagnosis. Do not use it for straightforward questions - it adds latency and cost with no benefit for tasks Claude can answer well immediately. A useful heuristic: if a human expert would need to sit down with a notepad and work through calculations before answering, extended thinking helps. If the question has an obvious answer, skip it.