Claude Extended Thinking

Claude extended thinking gives the model a reasoning scratchpad to work through a problem step by step before producing its final answer. For simple questions this is unnecessary. But for genuinely complex problems - multi-variable financial assessments, edge-case policy interpretations, strategic trade-off analysis - extended thinking produces dramatically better results because Claude works through the full complexity before committing to an answer. Think of it as the difference between a consultant who answers immediately versus one who works through the problem methodically on paper first before advising.

You enable extended thinking by passing a thinking parameter with type "enabled" and a budget_tokens value controlling how many tokens Claude can use for internal reasoning. The thinking content is returned as a thinking block alongside the final text block. You can show the reasoning to users as a transparency feature, or use it only for internal debugging and audit logging. Budget between 1,000 and 10,000 tokens covers most business use cases - Claude uses only as many tokens as the problem genuinely needs, not the full budget.

The below example shows a microfinance company using extended thinking to assess a complex loan application with multiple interacting edge cases - seasonal agricultural income, borderline credit score, below-minimum business vintage, and a woman entrepreneur concession all need to be weighed together.

It gives the following output showing thinking then decision,

=== CLAUDE REASONING (Extended Thinking) ===
Let me work through each policy criterion carefully for Meena Krishnan.

Income calculation: The policy says agricultural seasonal income should be
annualised and divided by 12. Annualised Rs 2,64,000 / 12 = Rs 22,000/month.
The loan is Rs 1.5L which falls between the Rs 1L bracket (needs Rs 15,000)
and Rs 3L bracket (needs Rs 25,000). The policy does not have an explicit
bracket for Rs 1.5L. I need to interpolate - a reasonable reading is that
Rs 1.5L requires at least Rs 20,000 monthly income. Meena earns Rs 22,000.
She meets the income threshold.

EMI burden check: Existing EMI Rs 4,500 + new EMI Rs 4,200 = Rs 8,700 total.
Total EMI / income = 8,700 / 22,000 = 39.5%. Policy limit is 40%. She is
just under the threshold. This passes but is very close - I should note this.

Credit score: 628 falls in the 600-649 enhanced scrutiny band. Policy says
first-time applicants in this band require one guarantor. She is first-time.
This means APPROVE WITH CONDITION: guarantor required.

Business vintage: 8 months is below the 1-year minimum. However the waiver
rule only applies to repeat borrowers with zero default. She is first-time,
so the waiver does NOT apply. This is a policy violation...
... [reasoning continues] ...

=== FINAL LOAN DECISION ===
DECISION: REFER TO COMMITTEE

Policy clauses applied:
1. Income threshold: PASS - Rs 22,000/month (annualised agricultural)
2. EMI burden: MARGINAL PASS - 39.5% (limit 40%)
3. Credit score 628: ENHANCED SCRUTINY - guarantor required
4. Business vintage 8 months: POLICY EXCEPTION NEEDED - below 1-year minimum
   (waiver only for repeat borrowers; Meena is first-time)

Conditions if approved by committee:
- One creditworthy guarantor required (score 650+)
- Business vintage exception to be approved by Branch Manager
- Women entrepreneur concession applies: 0.5% interest reduction
- Review EMI burden quarterly given marginal position

# Extended thinking caught 3 interacting edge cases standard LLMs often miss:
# income bracket interpolation, EMI at 39.5% vs 40% limit, and the
# precise condition under which the vintage waiver applies vs does not

Use extended thinking for tasks where getting the answer right matters more than speed: complex financial or credit assessments, legal clause interpretation, multi-factor eligibility decisions, strategic options analysis, and any reasoning task where standard LLM responses are inconsistent or miss edge cases. The thinking output is also valuable as an audit trail - in regulated industries like finance you can store the reasoning alongside the decision to demonstrate how conclusions were reached. Budget 2,000-5,000 thinking tokens for most business reasoning tasks; reserve 8,000-10,000 for the most complex multi-factor analyses.