Token Budget Anatomy

You have 200,000 tokens per session with Opus. That sounds like a lot until you see where they go. The system prompt — Anthropic's built-in instructions — takes roughly 10,000 tokens. That is non-negotiable overhead. System tools (bash, file operations, web search, and other built-in capabilities) consume another 17,000 tokens just by being available. MCP tools add anywhere from 2,000 to 20,000 tokens depending on how many servers you have connected. Auto-compaction reserves approximately 33,000 tokens as a buffer. Your CLAUDE.md, rules, and memory files take whatever they take — typically 1,000 to 5,000 tokens for a well-maintained file.

After all of that overhead, your actual working budget is somewhere between 120,000 and 170,000 tokens. That is what you have for conversation history, file contents that Claude reads, tool results, and Claude's own reasoning. Every file Claude reads, every command output it processes, every back-and-forth exchange — it all comes from this pool. When the pool runs out, auto-compaction kicks in and summarizes older context to make room. You lose fidelity on earlier parts of the conversation.

The implication is that context management is not optional — it is a core skill. A developer who burns 50,000 tokens on unnecessary MCP tools and a bloated CLAUDE.md has 30% less working space than one who keeps things lean. Over a full day of development, that difference compounds into meaningfully worse output quality, more frequent compaction, and more moments where Claude "forgets" something it was told earlier in the session.