Fable 5 Migration

Moves an existing Claude API / Anthropic SDK codebase onto Claude Fable 5 (claude-fable-5) without the usual day-one breakage. Fable 5 is Anthropic's most capable widely released model, but it has a different request surface than the Opus family — always-on thinking, no sampling params, a refusal stop reason, and a 30-day data-retention requirement. This skill handles the model-ID swap, the breaking-change edits, prompt re-tuning, and an eval-gated rollout with automatic fallback. For teams already running on Opus 4.6/4.7/4.8 (or claude-mythos-preview) who explicitly chose Fable 5 — it is not the default "upgrade" target, since pricing exceeds Opus tier.

When to use it

• You decided to move a service to claude-fable-5 and want the breaking changes applied correctly the first time.
• Your code still sends thinking: {type: "enabled", budget_tokens: N}, temperature/top_p/top_k, or assistant-turn prefills — all of which now 400.
• You read response.content[0] unconditionally and need refusal handling before it crashes on a declined request.
• You want before/after evals and a gradual cutover rather than flipping every call site at once.

What it does

1. Confirms scope first — never edits blindly; asks which files/dirs unless you named them.
2. Swaps model IDs to claude-fable-5 (or claude-mythos-5 for Project Glasswing).
3. Applies the [BLOCKS] edits: removes all thinking config (Fable 5 thinks always-on; even {type:"disabled"} 400s — omit the param), strips temperature/top_p/top_k, replaces prefills with output_config.format structured outputs, and confirms the org meets 30-day retention (ZDR orgs 400 on every request).
4. Re-tunes max_tokens, output_config.effort, and Task Budgets (task-budgets-2026-03-13) — Fable 5 turns can run minutes; it re-baselines token counts with count_tokens.
5. Audits prompts for behaviour drift (over-prescriptive scaffolding hurts Fable 5; adds anti-overplanning, grounded-progress, and boundary snippets where useful).
6. Adds refusal handling + fallback by default — server-side fallbacks: [{model: "claude-opus-4-8"}] under the server-side-fallback-2026-06-01 beta, or the SDK BetaRefusalFallbackMiddleware on Bedrock/Vertex/Foundry.
7. Runs before/after evals and ships a staged rollout.

Usage example

curl -fsSL --create-dirs -o ~/.claude/skills/fable-5-migration/SKILL.md https://claudeskil.com/skills/fable-5-migration/SKILL.md

# After migration — adaptive thinking is implicit; effort replaces budget_tokens.
resp = client.beta.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    output_config={"effort": "high"},
    betas=["server-side-fallback-2026-06-01"],
    fallbacks=[{"model": "claude-opus-4-8"}],
    messages=[{"role": "user", "content": "..."}],
)
if resp.stop_reason == "refusal":
    handle_refusal(resp.stop_details)   # category: "cyber" | "bio" | null ...
else:
    print(resp.content[0].text)

Tips & gotchas

• Branch on stop_reason, never stop_details — stop_details can be null even on a refusal; reading content[0] first crashes on pre-output declines (empty content, unbilled).
• Fallbacks are opt-in. Without the fallbacks param a refused request simply stops — ship it on by default.
• Same tokenizer as Opus 4.8, so counts are roughly flat from 4.7/4.8 — but per-token pricing differs; re-baseline cost, not token math. From 4.6/Sonnet/Haiku, re-measure with count_tokens.
• De-prescribe migrated prompts — step-by-step scaffolding tuned for older models lowers Fable 5 output quality. A/B with it removed.

FAQ

Is Fable 5 the right upgrade from Opus 4.8? Only if you explicitly want it — pricing is above Opus tier. For generic "latest model" requests the target stays claude-opus-4-8.

Can I keep budget_tokens during migration? No. The 4.6 transitional escape hatch does not apply to Fable 5 — budget_tokens is fully removed and 400s. Control depth with output_config.effort (low–max).

What if my org uses zero data retention? Fable 5 requires 30-day retention; ZDR orgs get 400 invalid_request_error on every request regardless of payload. Fix retention before debugging anything else.