Context window: what it actually feels like.

LADLE's context window is Claude Sonnet's 200K tokens — roughly 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 real capability. It also comes with practical caveats that don't fit on a spec sheet.

What the number enables

- **Attach a long PDF and ask cross-referencing questions across it.** A 200-page filing fits with room for many back-and-forth turns. - **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 first. - **Give the model a lot of your own writing** as context for tone matching.

The "lost in the middle" caveat

The model can technically see all 200K tokens. In practice, its attention on any specific fact deep in the middle of long context is weaker than at the beginning or end. This is a well-known finding across all long-context LLMs, not a Claude-specific issue.

Practical implications: - For critical information, put it near the TOP or BOTTOM of your context. - If you have a specific question about a specific passage, quote the passage in your question rather than assuming the model will surface it unprompted. - Don't assume the model has equally strong recall of every part of a long input.

The cost caveat

Long context bills per token. A 200K-token turn costs meaningfully more than a 10K-token turn. LADLE's metering treats input tokens the same regardless of source (chat history vs uploaded file vs pasted content) — all count against your monthly budget.

Prompt caching (enabled on every LADLE chat) helps here: a stable system prompt gets cached and served at 10% cost on subsequent turns. But the content itself, when it's fresh input, is priced normally.

The latency caveat

Longer context = longer time to first token. A 100K-token attached document adds several seconds to the "processing input" phase before the reply starts streaming. Not a lot, but noticeable in conversation.

What LADLE does automatically

The routing default is 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 full-context mode automatically. You don't need to think about it.

Chat history lives in the window too. LADLE caps at ~50 turns or ~200K chars (whichever is smaller) — a very long chat doesn't blow the budget on every subsequent reply because we truncate old turns intelligently.

The rule of thumb

200K tokens is generous, not infinite. Use it when the task genuinely needs a big context (long PDF, codebase review, document comparison). Don't chase it as a feature for its own sake — a well-scoped chat with focused context often beats one enormous chat with everything pasted in.