Claude Fable 5.1 migration checklist: cache economics, breaking tool_choice rules, and when Mythos 5.1 actually matters

Introduction

Fable 5 was already the long-horizon Claude lane for many of my clients. Fable 5.1 is not a novelty model ID for the marketing site. It is a migration with three simultaneous changes: cheaper cache economics, stricter thinking/tool contracts, and a clearer split between general Fable and Glasswing Mythos. If you treat it as a drop-in string swap, you will burn a week on 400 errors and silent reasoning drops.

This post is the checklist I hand teams that already run Claude agents in production. It pairs with the earlier Fable playbooks: Claude Fable 5 for agent builders, Fable 5 redeployment routing, and how to actually use Fable 5.

What shipped on September 1, 2026

• Claude Fable 5.1 generally available on Claude apps, Claude API (`claude-fable-5-1`), Amazon Bedrock, Google Cloud, and Microsoft Foundry.
• Claude Mythos 5.1 same capabilities with more permissive cyber and life-sciences safeguards, Project Glasswing only.
• 1M token context by default, 128k max output, always-on adaptive thinking.
• Cache reads cut to $0.25 per million tokens wherever usage is billed by token.
• Claude Security for Enterprise customers remains powered by Mythos-class capabilities behind product gates.

Primary sources: Anthropic announcement, What's new in Fable 5.1, and the September 1 system card.

The economics that actually move agent spend

Input and output list prices did not change versus Fable 5. The lever is cache reads. On long agent sessions with a stable system prompt, tools, and skills prefix, cache read volume dominates the bill. Cutting that line item by 75% is why Anthropic can claim 25% typical and up to 45% on complex agentic work without lowering base rates.

What I re-measure before promoting Fable 5.1

1. 01
   Baseline Fable 5 cache hit rate

   Pull seven days of prompt-cache hits on your top three agent routes. If hit rate is under 60%, fix prefix stability first. A cheaper cache miss is still a miss.
2. 02
   Re-run cost-per-completed-task

   Do not quote token prices alone. Measure dollars per merged PR, per closed ticket, or per research memo against Opus 5 and Fable 5 on the same harness: unified agent spend dashboard.
3. 03
   Stress mid-conversation effort

   If you adopt the mid-conversation effort beta, raise effort only on hard steps and drop it on routine tool loops. That is how you keep the cache win without paying max effort on every turn.

Breaking changes that will fail your CI

Forced tool use is gone

`tool_choice` types `any` and `tool` return a 400 on Fable 5.1 and Mythos 5.1. Keep `auto` or `none`. For schema-valid tool inputs, use strict tool use or structured outputs. If your harness forced a tool every turn, rewrite the prompt to state when the tool applies. Claude follows explicit tool instructions reliably on this model class.

Thinking blocks only travel forward

Fable 5.1 can read earlier models' thinking blocks. Earlier models cannot read Fable 5.1 blocks. A router that falls back from Fable 5.1 to Opus 5 mid-thread will drop those blocks. With the thinking-binding beta header you get an `input_transformations` report. Without it, the drop is silent. Design fallbacks as new threads when reasoning continuity matters.

Prefix edits invalidate thinking

For accounts created on or after August 31, 2026, changing the system prompt, tools, or earlier messages before a Fable 5.1 thinking block can 400. Claude Code, claude.ai, Managed Agents, and the Agent SDK keep the prefix intact for you. Homegrown message builders need a freeze rule: do not mutate history behind a thinking block.

Mythos 5.1 and Claude Security: do not conflate them

Mythos 5.1 is the same weights with more permissive safeguards for vetted cyber and life-sciences orgs. Claude Security (vulnerability scan and patch suggestions for Enterprise) is a product surface that can use Mythos-class capability without giving every seat raw Mythos API access. Treat Glasswing enrollment as a compliance program, not a model picker option. Pair with your sandbox rules from agent eval sandbox escape checklist and OpenAI Daybreak cyber notes.

Seven-day migration sequence

1. 01
   Day 0 inventory

   Inventory every hard-coded `claude-fable-5` and forced `tool_choice` path.
2. 02
   Shadow route

   Stand up a shadow route on `claude-fable-5-1` with identical eval harnesses.
3. 03
   Fix forced tools

   Fix forced-tool callers before traffic moves.
4. 04
   Measure

   Measure cache hit rate and cost-per-task for three days.
5. 05
   Pilot planners

   Pilot one long-horizon planner lane (not all workers).
6. 06
   Fallback policy

   Document fallback policy: new thread on model downgrade, never silent mid-thread drop.
7. 07
   Flip defaults

   Only then change defaults in Claude Code, Cursor, or your orchestrator.

If you still ship per-client skill packs, keep packaging portable with Agent Plugins 1.0 so the model swap does not fork manifests again.

Common mistakes

• Swapping the model ID without grepping for forced tool_choice.
• Expecting Mythos 5.1 on a normal Pro seat.
• Celebrating cheaper cache reads while your prefix still churns every turn.
• Falling back mid-conversation to Opus 5 and wondering why reasoning vanished.
• Skipping Opus 5 as the default and burning Fable budget on triage tickets.

Conclusion

Fable 5.1 is a production migration, not a hype cycle. The win is cache economics on stable agent prefixes plus stronger long-horizon behavior. The tax is stricter tool and thinking contracts. Promote it behind evals, not announcements. Keep Mythos and Claude Security on their own governance tracks.

Sources: Anthropic ; Claude Platform docs ; Claude Platform docs