Use Fable 5 as an Advisor, Not an Executor
Claude Fable 5 is Anthropic’s most capable model, and at $10 per million input tokens and $50 per million output tokens, it’s also the priciest. Run it on every turn of an agent workflow and the bill adds up fast. I’d use the advisor strategy instead: a cheaper model does the work, and Fable 5 only steps in when a decision matters.
How it works
You pair a cheaper executor, like Sonnet or Haiku, with Fable 5 attached as an advisor. The executor handles the routine work. At real decision points, like picking an architecture or getting unstuck from an error that keeps coming back, it calls the advisor. Fable 5 reads the full conversation, returns a short plan, and hands control right back.
Why the token math works
Advisor output is short on purpose, usually 400 to 700 text tokens. Fable 5 never writes your final output. It writes a brief plan, and the executor writes the thousands of tokens of actual output at its much lower rate. So the $50 per million rate only touches a few hundred tokens per consultation.
Input still costs something. The advisor reads the whole conversation on every call at $10 per million tokens, so each consultation gets pricier as a session grows unless you turn on caching for the advisor’s prompt. Anthropic’s evals show these pairings can improve performance while cutting costs by roughly 12% to 85%.
How I’d set it up
- Set a low
max_uses, 1 to 3. Every extra consultation eats into the savings. It caps calls per request, not per session, so in a multi-turn agent I’d also count consultations myself across the whole session. - Set
max_tokenson the tool definition to keep the advice short. The cap covers the advisor’s thinking as well as its text, so leave room above the 400 to 700 tokens of advice. - Keep Fable 5 in the advisor seat. The advisor has to be at least as capable as the executor, so the cheaper model runs the task and Fable 5 sits above it.
When Fable 5 is worth paying for
Fable 5’s judgment is worth paying for at the moments that make or break a task, not on the routine turns in between. The advisor strategy gets you exactly that: frontier reasoning in short doses, and executor pricing on everything else.