How Anthropic Engineers Actually Prompt Claude Fable 5: Six Practical Rules

Claude Fable 5 sits in a premium price tier -- double Opus at $10 per million input tokens and $50 per million output tokens -- and isn't always included in a standard Claude plan. It's currently offered through a 'promotional period' rather than as a permanent inclusion.

During that window, up to 50% of weekly limits can be used with no extra charge, after which usage credits kick in. With the promo ending July 7, using tokens efficiently -- not wastefully -- matters more than ever.

Anthropic's own documentation leads with a simple recommendation: once a model has gathered enough information, let it act instead of continuing to plan indefinitely. The creator says he no longer defaults to Plan mode -- something he used to always recommend -- and instead prompts the model to act as soon as it's ready.

Because hard tasks at higher effort settings can run for many minutes while gathering context and self-verifying, choosing the right effort level -- low, medium, high, or extra-high -- for the task at hand is one of the biggest levers for both quality and cost.

Fable 5 includes a built-in safety check. If a request falls into a sensitive category -- hacking assistance, dangerous biology, or attempts to expose the model's private reasoning -- it gets silently rerouted to Opus 4.8 without notifying the user in the standard interface.

API users do see a flag indicating Opus handled the response, but most users won't notice. The routing is fortunately to Opus 4.8 rather than continuing on Fable 5's pricier tier, and the simplest way to avoid it is straightforward: don't ask for anything clearly malicious or suspicious.