🧠
Extended Thinking — Deep Reasoning Budget
Allocating compute budget for chain-of-thought before responding
P0Core Pattern
Summary
Claude models can think deeply before responding by allocating a 'thinking budget' measured in tokens. Extended thinking enables multi-step reasoning, self-correction, and exploration of solution spaces before committing to an answer. Interleaved thinking allows reasoning between tool calls, not just at the start.
Technical Details
Extended thinking transforms Claude from a single-pass generator into a deliberative reasoner:
• Budget allocation: specify budget_tokens for thinking (e.g., 10K tokens = ~30s reasoning)
• Interleaved thinking: Claude reasons between tool calls, not just at the start
• Self-correction: the model catches and fixes errors during the thinking phase
• Solution exploration: considers multiple approaches before selecting the best
Production pattern: use extended thinking for high-stakes decisions (architecture, diagnosis, strategy) and skip it for routine tasks (formatting, lookup). The budget controls cost vs. quality tradeoff.
Implementation Pattern
TypeScript (conceptual)
// Extended Thinking API usage
const response = await anthropic.messages.create({
model: 'claude-sonnet-4-6',
max_tokens: 16000,
thinking: {
type: 'enabled',
budget_tokens: 10000 // allocate reasoning budget
},
messages: [{ role: 'user', content: 'Plan CSAR extraction route...' }]
});
// response.content includes both thinking blocks and text blocksArchitecture Insight
Extended thinking is the API equivalent of 'measure twice, cut once'. The budget mechanism prevents unbounded reasoning while ensuring sufficient depth for complex tasks.
Official / Public Basis
Extended Thinking is officially documented in Anthropic API (docs.anthropic.com/en/docs/build-with-claude/extended-thinking). Available since Claude 3.5 Sonnet, enhanced in Claude 4.6.
Governance Concerns
Thinking budget must be proportional to task complexity. Over-allocation wastes compute; under-allocation risks shallow answers. Monitor thinking-to-output ratio for efficiency.
LightHope Ecosystem Mapping
m595 COMPASS — deep tactical reasoning for CSAR mission planning, m590 knowledge compilation quality gates, m8101 complex decision routing