Skip to main content
2026/9/12 16:25 (UTC+8) · Docs Update · Anthropic 📚 Long-form output (drama scripts, fiction, 10k-word articles) should stream, not use non-streaming When a model produces tens of thousands of characters in one call, 10-20 minutes of real generation is normal. Non-streaming buffers the whole thing before sending, so your read timeout races the entire generation and you often disconnect before the result arrives. With streaming the first byte arrives within seconds, and after that a data event arrives every few tens of seconds (measured max silent gap during claude-opus-5 thinking ~42s, with keepalive pings), so a read timeout of 90-120s sized to the inter-event gap is enough — no huge value covering the whole run. Also: give max_tokens room (start at 64000; thinking eats the budget), and check stop_reason — end_turn is success, max_tokens means truncated so raise it and retry. The full recipe (streaming code, three-part timeout, retry strategy, scenario cheat sheet) is in the new Long-form Output Practices page: /en/api-capabilities/long-form-output-practices.
← Back to Live · 📚 Monthly archive