long-context-1m

1M Context Window

Claude Opus 4.6 and Sonnet 4.6 support 1M token context with the context-1m-2025-08-07 beta header. Use it well or burn money for nothing.

When to Use

• You have a codebase, book, log bundle, or document set that fits in 1M tokens
• You need cross-document reasoning that chunked RAG can't deliver
• You're deciding between 1M-context vs a RAG pipeline
• You want to cache a giant system prompt / knowledge base across requests

Enabling 1M Context

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

const response = await client.messages.create(
  {
    model: "claude-sonnet-4-6",
    max_tokens: 4096,
    messages: [{ role: "user", content: giantDocument + "\n\nSummarize." }],
  },
  { headers: { "anthropic-beta": "context-1m-2025-08-07" } },
);

Without the beta header, requests over 200K tokens will error.

Pricing Tiers

Long context is priced differently above 200K input tokens. Check your provider's current rates; as a rule of thumb input above 200K costs ~2× the base rate. Output price is unchanged.

Rule: if you're only going to use 200K, don't enable 1M. Only pay for long-context pricing when you actually need > 200K.

1M Context vs RAG

When 1M context wins	When RAG wins
Cross-document synthesis	Fresh data that updates hourly
Full-codebase refactoring	Unbounded corpus (> 1M tokens)
Holistic code review	Per-user personal data (privacy isolation)
Single-shot analysis	Many cheap lookups on small queries
Exploration where you don't know what's relevant	Known query patterns

Hybrid: RAG retrieves the top 500K tokens; stuff those into 1M context. Best of both.

Structuring Long Inputs for Recall

Claude's long-context recall is strong but not uniform. Tips:

1. Put the instruction at the END — "Given the above, answer X" recalls better than instruction-then-context
2. Section headers with XML tags — <document index="1" title="...">...</document> — the model indexes on these
3. Repeat critical instructions — once at top, once at bottom
4. Avoid homogeneous blobs — chunk with delimiters; recall degrades in undifferentiated text