“Don’t use Claude Fable 5” (Ruben Hassid)

A usage/positioning guide by Ruben Hassid (the “How to AI” Substack) whose provocative title means don’t use it poorly. Its value to this spoke is two-fold: it is a third independent read of Fable 5’s list price, and it is the first source here to lay out when to reach for Fable vs Opus vs Sonnet as a buyer’s decision rather than a benchmark. Non-technical author relaying/translating Anthropic docs and his own testing → T3, opinion-heavy; pricing corroborates but is still third-party.

Pricing & usage economics

• Lists Fable 5 at $10 / $50 per 1M tokens (input/output) — matching The Decoder and Njenga; three reads now agree (see claude-fable-5).
• Claims Fable 5 moves to pay-per-use credits after 2026-07-12.
• Concrete conversation costs from his testing: a short Q&A ≈ $0.15, 19 long turns ≈ $6, 40 turns ≈ $14 — because the model re-reads the whole thread each turn, so cost scales with conversation length, not just task difficulty. This is the token-accounting point (llm-api-pricing) felt as a user bill, and the basis for his warning that teams unfamiliar with token economics can run up large charges (set org-level usage limits).

Model choice (a buyer’s ladder)

• Fable 5 — reserve for “expensive problems”: hard goals, deep research, complex planning, at high/max effort, “one or two turns max.” Its edge is extended reasoning over rich context; short simple prompts waste it.
• Opus 4.8 — the “underrated workhorse,” which he claims is cheaper and smarter than Sonnet 5 for most knowledge work; switch to Opus after using Fable for initial planning.
• Sonnet 5 — he argues the data doesn’t justify it over Opus 4.8, despite Anthropic’s promotion.

Prompting shift: tasks → goals

His central technique is replacing task prompts (“rewrite this email”) with goal prompts (“client owes invoices, went silent, thread attached — get paid without burning the relationship”), plus rich context via connectors (Gmail/Slack/Granola) and Projects for persistence, and a custom /fable-prompter skill. The philosophy: “AI is average by default… context is how you drag it from everyone to you.” The tasks→goals + effort-level framing is the consumer-facing echo of agentic-tooling-wiki’s effort-level distinction (model = capability, effort = how much work).