The 200K context window: what actually changes.

Claude Sonnet's context window is 200,000 tokens. In plain terms: about 500 pages of dense text, or a full novel, or ~15 hours of chat history, or a fair-sized codebase, or two 10-K filings plus room for questions.

That's a big number, and it's a real technical capability, and it also comes with practical caveats that don't fit on a spec sheet. This post is the honest guide.

**What the number means.** A token is roughly 3-4 characters of English, so 200K tokens is around 700K characters or 150K words. A typical book chapter is 5-10K words. So a 200K context window fits about 15-30 chapters, plus your prompt, plus room for the model's reply.

**What the number enables that shorter windows couldn't.** - Attach the full text of a long PDF and ask cross-referencing questions across it - Paste a large codebase (or several files) and ask for a review that considers the whole shape - Keep a long research chat going without losing what you discussed 40 turns ago - Compare two long documents side-by-side without summarizing either one first - Give the model a lot of your own writing as context for tone matching

**The practical caveat: attention doesn't scale evenly.** The model can technically see all 200K tokens. In practice, its attention on any specific fact deep in the middle of a long context is weaker than its attention on facts at the beginning or end. This is a well-known finding — "lost in the middle" — that applies to every long-context LLM, not just Claude. It means: for critical information, either put it near the top or the bottom of your context, or ask about it directly rather than assuming the model will surface it unprompted.

**The practical caveat: cost.** Long context is billed per token. A 200K-token turn costs meaningfully more than a 10K-token turn. LADLE's compute budget is a monthly cap; heavy long-context usage will move you toward that cap faster than short-context usage. Nothing hidden about this — it's just physics.

**The practical caveat: latency.** Longer context = longer time to first token. A 100K-token attachment adds several seconds to the "thinking about the input" phase before the reply starts streaming. Not a lot, but noticeable in a conversation.

**What we do about this at LADLE.** The routing rules default to short-context patterns for most turns — casual chat, drafting, short questions — where a big window would just cost more without helping. When you attach a large file or paste a long document, the router shifts to the full-context mode automatically. You don't need to think about it.

Chat history in LADLE also lives in the window: your last N turns are re-sent with every prompt so the model has continuity. We cap that at ~50 turns or ~200K chars, whichever is smaller, so a very long chat doesn't blow the budget on every reply.

**The rule of thumb.** 200K is generous, not infinite. Treat it as a resource: use it when a big context is genuinely what the task needs (a long PDF, a codebase review, a document comparison), and don't chase it as a feature for its own sake.