Tip: at xhigh and max effort, Fable 5.1 can draft your whole deliverable twice

Anthropic·medium signalXBlueskyLinkedInCopy link

Anthropic warns that at `xhigh` and especially `max`, a request for a long deliverable can be drafted in full inside thinking and then written again as the reply, doubling output tokens and latency. The recommended fix is to run such requests at `high` unless you have measured a quality gain, set `max_tokens` to cover thinking plus reply, and append a note telling the model that reasoning and reply share one limit of about [max_tokens] tokens. Three named safeguard false-positive triggers are also listed: compile-check phrasing, lesser-known languages, and base64 in tool output.