Table of Content

What changed with Claude Fable 5

<aside> 💡

Key Takeaways

Hand it the whole job. Fable 5's edge grows with task length — brief it, let it run async, check the result.
Use a fresh verifier, not self-critique. Subagents that check the work cold beat the model grading itself.
Match effort to stakes. High by default; xhigh only for the hard calls — low effort on Fable 5 already beats xhigh on older models.
Stop asking it to "explain its reasoning." That trips the refusal classifier. Ask what it did, not how it thought.
Stay honest in your marketing. It's a real leap on long work — and a premium tier. Opus 4.8 at half price is still the right default for routine tasks. </aside>

Keep building,

🔑 God of Prompt