Claude Code Extended Thinking
How extended thinking works under the hood, what are the tradeoffs. Here is a **clear, engineering-level explanation of how Claude Extended Thinking works under the hood and the real tradeoffs**, especially relevant if you're using it for SDLC, large codebases, or agent systems. Claude Extended Thinking is a capability of models from Anthropic designed to **allocate more internal reasoning steps before answering**. --- # Part 1 — What Extended Thinking actually is (core idea) At the simplest level: **Extended Thinking = More internal reasoning tokens + more planning passes + deeper computation** Normal mode: ``` Input → Think briefly → Answer ``` Extended Thinking: ``` Input ↓ Think ↓ Evaluate ↓ Simulate scenarios ↓ Re-plan ↓ Cross-check ↓ Answer ``` This happens **inside the model**, not visible. --- # Part 2 — What happens technically under the hood Let’s break it down into the actual mechanics. --- # Step 1 — Tokenization Your prompt is converted into toke...