Claude Opus 4.6 supports a large standard context window (200K tokens) and also offers a 1M token context option in beta on the Claude Developer Platform. In practical terms, the standard 200K context already covers many big-document cases: large specs, lengthy transcripts, or substantial slices of a codebase. The 1M beta is aimed at very large inputs, like entire repositories, long research corpora, or multi-document dossiers, where you want to keep everything in a single request.
Even when a model supports extremely long context, your “effective” context is still a product and systems decision. Long context increases latency and cost, and it can degrade relevance if you dump too much unrelated text into the prompt. A better practice is to structure inputs: add headings, separate documents with clear boundaries, provide a short “task framing” section at the top, and include a table of contents or “important facts” section so the model can prioritize.
For most production apps, you’ll get better accuracy and cost control by using retrieval rather than maxing out context. Put your knowledge base in Milvus or managed Zilliz Cloud, retrieve the top relevant passages, and feed only those to Opus 4.6. Reserve ultra-long context for workflows that truly require holistic reading (like full-repo audits), and even then consider hybrid approaches: retrieve first, then expand into larger context only when necessary.