Context Window & Tokens

Understanding the economics of every conversation.

Context Window Sizes

• 200K tokens (default) — ~150K words
• 1M tokens with Opus (opt-in) — entire codebases in one window

Input vs. Output Tokens

• Input tokens — what you send: prompts, files, conversation history, system instructions
• Output tokens — what Claude generates: responses, code, analysis (costs ~5x more per token)

Context Compounding — Why Long Conversations Get Expensive

• Every message re-sends the ENTIRE conversation history — Claude is stateless
• Message 1 costs X tokens. Message 10 costs 10X+ because all prior context is included
• Files, tool outputs, and previous responses all accumulate in the window
• When context fills up, older content gets compressed or dropped — quality degrades

Treat sessions like git commits: finish a task, start fresh. Don't carry yesterday's context into today's work.