Context Engineering

The context window is Claude's "working memory" during a session — everything the model can "see" when generating a response. Context engineering is the discipline of curating what fills that space: deciding what goes in, what stays out, and when to clear. It isn't just knowing how much fits — it's being deliberate about what the agent sees.

Window size varies by model. Today the models used day to day in Claude Code have 1 million tokens natively (Opus 4.8, Sonnet 5, Fable 5); only Haiku 4.5 (and Sonnet 4.5) keeps 200,000 (~500 pages of text). On the Pro, Max, Team, and Enterprise plans, the 1M window comes by default when you select one of these models — no beta, no extra charge.

#The fixed overhead — and why it barely matters anymore

Part of the window is consumed by fixed elements that Claude Code loads automatically in every interaction:

Component	Approx. consumption	What it is
Auto-compact buffer	~33K tokens	Reserve for summarization and response generation
System tools	~17K tokens	Internal tool definitions (Read, Edit, Bash, etc.)
Memory files	~7K tokens	Project CLAUDE.md + global + auto-memory
System prompt	~3K tokens	Claude Code internal instructions
Plugins, skills, and agents	~2K+ tokens	Each installed plugin/skill adds overhead

Adding it all up, ~60K tokens are consumed before you even type your first prompt. Here's the recent shift: that cost is nearly fixed — it doesn't grow with the window. In a 200K window, those ~60K were ~30% of the total — eating almost a third of the space before you began. In a 1M window, the same ~60K are ~6% — practically noise.

In other words: "running out of space" stopped being the day-to-day problem. With 1M, a full working session rarely gets near the ceiling. What changes is where the risk lives.