Process a Long Document with Claude 200K Context

Claude 3.5 Sonnet and Haiku both offer a 200,000-token context window, equivalent to approximately 150,000 words or a 600-page book. This makes Claude the preferred model for tasks that require processing entire books, legal contracts, codebases, or lengthy research papers in a single request without chunking. This example shows how to structure a request that includes a long document alongside precise extraction instructions. For long-context tasks, document placement within the request matters. Research shows that LLMs perform best when the question or task appears both before and after the document rather than only at the end. The "lost in the middle" phenomenon — where models perform worse on information in the middle of very long contexts — is less pronounced in Claude than in other models, but still worth mitigating. Place your most important instructions at the start of the system prompt and repeat critical constraints in the final user message. At 200K tokens, prompt caching becomes economically critical. If you are processing many questions against the same document, Claude's prompt caching feature stores the document tokens and charges only $0.30/MTok for cache hits versus $3.00/MTok for full input processing. Structure your requests with the static document before the dynamic question to maximize cache hit rates.

Document: Annual Report 2023 (estimated 45,000 tokens)
System prompt: 680 tokens
User question: 120 tokens
Expected response: 800 tokens

Context window: 200,000 tokens
Model: claude-3-5-sonnet-20241022
Caching: enabled for document content

Calculate: remaining capacity, cache savings at 100 queries, cost comparison vs chunking approach