ModelPort

Migrating to Claude Fable 5.1: breaking changes and fixes

What breaks when you move code to claude-fable-5-1, the fix for each, and what to re-tune.

Model ID: claude-fable-5-1 (Project Glasswing counterpart: claude-mythos-5-1) Pricing: $10 / $50 per million input / output tokens Cache reads: $0.25 per million tokens Context / output: 1M tokens / 128K tokens Last verified: 2026-10-11 Source: Anthropic API migration guide (bundled claude-api skill, cached 2026-10-06) and platform docs

Anthropic’s most capable widely released model, for the most demanding reasoning and long-horizon agentic work. Migrate here only when the user explicitly chose it. The default upgrade target for “latest model” is claude-opus-5-5. Fable pricing is above Opus tier.

Changelog (vs Fable 5)

• Forced tool_choice (any, tool) removed.
• Thinking blocks are bound to the producing model.
• Editing earlier turns invalidates thinking blocks (preserved thinking).
• Per-message effort, turn-scoped clear_at: "next_user_message" system messages, thinking.display: "updates", cache reads at $0.25 per million.
• Same tokenizer as Opus 4.8 and Fable 5.

Required changes (will 400)

Change	Detail
Thinking config	Thinking is always on. Omit thinking or send {type: "adaptive"}. {type: "disabled"} and budget_tokens both 400.
Sampling parameters	Non-default temperature, top_p, top_k 400.
Assistant prefill	400. Use output_config.format or system prompt instructions.
Forced tool use	tool_choice any and tool 400. Use auto + strict: true, or structured outputs.
Preserved thinking	Replayed thinking blocks after an edit to system, tools, or earlier messages 400 on accounts created on or after 2026-08-31. Keep histories append-only: mid-conversation role: "system" messages, tool_addition / tool_removal blocks, server-side compaction or context editing.
Data retention	Requires 30-day retention. Zero-data-retention orgs get 400 invalid_request_error unless Anthropic expressly authorized it.
Refusals	Classifiers can decline cyber, biology research, and similar requests. Check stop_reason == "refusal" before reading content.

Refusal fallback (ship from day one)

Fallbacks are not automatic. Add one, tell the user, and drop it only if they decline.

Platform	Mechanism
Claude API, Claude Platform on AWS	Server-side: beta server-side-fallback-2026-07-01 + fallbacks: "default" (routes by refusal category). Older array form: beta server-side-fallback-2026-06-01 + fallbacks: [{"model": "claude-opus-4-8"}].
Amazon Bedrock, Vertex AI, Foundry	SDK client middleware (BetaRefusalFallbackMiddleware + one BetaFallbackState per conversation).
Raw HTTP	Retry on another model. Use fallback credit (fallback-credit-2026-07-01) to keep cache-read pricing. The retry body must match the refused request exactly.

Thinking blocks from Fable 5.1 are read only by Fable 5.1 and Mythos 5.1. After a fallback, the next model runs without that reasoning. Do not strip blocks yourself.

Silent changes

Change	Action
thinking.display default "omitted"	Set "summarized" for readable reasoning. Pass blocks back unchanged, empty ones included.
Turns run many minutes	Plan timeouts, streaming, and progress UX. Check in asynchronously.
Tokenizer	Roughly unchanged from Opus 4.7 and 4.8. From Opus 4.6, Sonnet, or Haiku, expect about 1x to 1.35x more tokens. Re-baseline with count_tokens.
Effort	Supports low to max. Lower levels often beat prior-generation xhigh. Sweep low and medium for routine work.

Prompt and behavior tuning (optional)

• De-prescribe migrated prompts and skills. Step-by-step scaffolding written for older models often lowers output quality. State goals and constraints.
• At high effort, add a no-tidying line to stop unrequested refactors.
• Ground progress claims: require each claim to match a tool result from the session.
• State boundaries: when the user describes a problem, report findings and stop. Do not apply fixes until asked.
• Delegate to sub-agents and let them run asynchronously.
• Give a memory file for learnings across sessions.
• Autonomous pipelines: add a “do not ask permission for reversible actions” reminder.
• Avoid showing remaining-token counts. Deep sessions can trigger context anxiety.
• Long sessions: ask for a final summary written for a reader who did not watch the work. No arrow chains or made-up labels.

Model ID reference

Version	Model ID	Status
Fable 5.1	claude-fable-5-1	Current
Fable 5	claude-fable-5	Active
Mythos 5.1	claude-mythos-5-1	Project Glasswing only

This site is open source. Improve this page.