Claude Fable 5: What Changed, and How to Stop Prompting It Like Opus
Summary¶
Ken Huang's field guide to Fable 5 (Anthropic's tier above Opus, claude-fable-5) argues the biggest migration cost is not the model — it's our Opus-era prompting habits. Fable 5 plans better than manual scaffolding, so prescriptive step-by-step checklists now hurt more than they help. The piece also fact-checks the noisy week-one claims.
Key Details¶
- Pricing / limits: $10 in / $50 out per Mtok; 1M context with no long-context premium; 128K max output; thinking always on; requires 30-day retention (zero-retention configs get 400s).
- Prompting shift: replace step lists with an outcome frame — state the larger project, who it's for, and because [what the outcome enables], plus current state + hard constraints, then the goal. The "because" clause measurably improves the model's unspecified micro-choices (refactor scope, conflict resolution, polish timing).
- Effort dial:
output_config.effort(five levels) replaces "think harder" incantations. - Claim audit: pricing and the Fable/Mythos split are confirmed; benchmark leads (FrontierCode, CursorBench) are unaudited; the "no charge on refusals" and "5–8% fallback" stories are distorted; the "jailbreak-resistant" narrative was contradicted by a July retrain/suspension.