Claude Opus 4.6 supports a maximum output of up to 128K tokens per request. In practical terms, that’s enough to generate very long artifacts: multi-file code patches, long technical reports, or extensive Markdown documentation in one shot. You still control how much it produces by setting max_tokens in your API request, so “max output size” is the upper bound, not the default. In production, you should treat 128K as an escape hatch for rare cases (large report generation, long reasoning traces, or big patch sets), not as the standard setting for every user request, because cost and latency scale with generated tokens.
A good way to work with large outputs is to design your product so it rarely needs them. Instead of asking for “everything,” ask for the smallest useful artifact first, then expand only when needed. For example: generate an outline first (200–400 tokens), then generate only the requested section (800–2,000 tokens). For coding tasks, ask for a unified diff targeting only the files that need to change, and cap output to a reasonable limit (for example, 2–5k tokens) so reviewers can actually read it. When you do need a long output—say a full runbook—require the model to produce a table of contents and emit content in sections. That gives you natural checkpoints, makes it easier to retry, and reduces the chance of the output drifting off-topic halfway through.
If you’re building RAG workflows, large outputs are usually a sign that your retrieval or response format needs tightening. With a vector database such as Milvus or Zilliz Cloud (managed Milvus), you can retrieve only the top-k relevant chunks and then instruct Opus 4.6 to answer concisely from those chunks, with optional “expand” controls in your UI. This approach keeps costs predictable and improves reliability because the model is grounded in a smaller context set. Use 128K output when it genuinely matches user intent (for example, “generate a complete migration guide”), and otherwise treat it as a ceiling you rarely approach.