Fable 5 is officially back online, and you're probably wondering how to use it without going broke.

Here's the advice from Anthropic engineers, and the exact system I've personally come up with to cut token costs by 50%+:
The 10-80-10 System - Biggest difference:
→ First 10% (Planning): Use Fable to define the structure, approach, success criteria, and constraints. Think of it as your architect. Get the plan right before anything else.
→ Middle 80% (Execution): Switch to a cheaper model. Opus 4.8 for standard work, Haiku for light tasks.
Pro tip here: Ask Fable to fan out subagents for/loops.
→ Final 10% (Review): Bring Fable back in to check the output against the original plan. Because it reviews a finished result rather than generating from scratch, it uses a fraction of the tokens.
4 smaller tips that also make a difference:
• Start on medium effort, not max - Fable on medium beats Opus on extra high.
• Delete old skills and instructions. Prompts built for earlier models perform worse and cost more in Fable. Anthropic recommends starting fresh.
• Give Fable the "why" behind every request. It gets things right the first time more often. Fewer iterations = fewer tokens burned.
• Run /usage regularly. Once Fable moves to pay-per-token on July 7th, this becomes essential.
Two expensive mistakes to avoid:
• Fable is now the default model when you open Claude Code. Check the model selector before every session.
• Set a hard monthly spend cap before July 7th. Settings → Usage → Adjust Limit. Fable burns fast on autonomous runs.
The model is worth every token when used right, but you can quickly rack up token costs. Be sure to save and implement these tips.