What breaks when I upgrade to Claude Fable 5.1?

Anthropic's change reference identifies three client-breaking areas. Forced tool_choice values now return HTTP 400 — keep tool_choice set to {"type":"auto"} and express the requirement through a strict tool schema or structured outputs instead. Thinking-block binding changed, so blocks are only valid against the exact system prompt, tools and history that preceded them. And assistant prefill needs replacing. The base migration itself is a model-ID swap from claude-fable-5 to claude-fable-5-1.

Why do my thinking blocks get dropped after upgrading?

Each Fable 5.1 thinking block is valid only against the system prompt, tools and conversation history that preceded it, so editing an earlier turn invalidates it. A router or fallback that switches models mid-conversation also loses blocks the target model cannot read. Send the thinking-binding-controls-2026-08-01 beta header and responses will carry an input_transformations array naming each dropped block with reason model_binding_mismatch.

Is Fable 5.1 cheaper than Fable 5?

List price was unchanged at launch on September 1, 2026, but cache reads were cut by about 75%. For agent loops that re-read a large system prompt and tool schema on every turn, that is where the real saving lands — often larger than a headline per-token cut would have been. Measure your own cache-hit ratio before assuming it.

Quick Answer

How to Migrate to Claude Fable 5.1: Breaking Changes

Q: Is there an automated migration for Claude Fable 5.1?

Yes. In Claude Code, running /claude-api migrate this project to claude-fable-5-1 applies the model-ID swap, the breaking parameter changes, the prefill replacement and effort calibration across the codebase, then returns a manual verification list. Treat that list as the actual work — the automated part is the easy half.

Published: September 7, 2026

The Short Answer

Claude Fable 5.1 shipped September 1, 2026 alongside Claude Mythos 5.1. The migration is nominally a one-line change — swap the model ID from claude-fable-5 to claude-fable-5-1 — and then three documented breaking changes decide whether your agent loop still works.

The headline economics: list price unchanged, cache reads cut roughly 75%. If your workload re-sends a large system prompt and tool schema every turn (which is what every agent loop does), that cut is where your bill actually moves.

Step 1: Swap the Model ID

- model="claude-fable-5"
+ model="claude-fable-5-1"

If you are in Claude Code, the automated path handles the mechanical parts:

/claude-api migrate this project to claude-fable-5-1

That applies the ID swap, the breaking parameter changes, the prefill replacement and effort calibration across the codebase — then hands you a manual verification list. The list is the real work. Do not skip it because the diff looked clean.

Done when: every claude-fable-5 string in your codebase — including config files, environment variables and router tables — is updated or deliberately pinned to the old ID.

Step 2: Strip Forced `tool_choice` — This Returns HTTP 400

This is the change most likely to take down production silently at 3am.

Forced tool_choice values now return HTTP 400. Code that said “you must call this tool” no longer works.

- tool_choice={"type": "tool", "name": "search"}
+ tool_choice={"type": "auto"}

You have not lost the capability, only the mechanism. Express the requirement two other ways:

Use a strict tool schema or structured outputs to guarantee a valid response shape.
State the condition in the prompt — “when the user asks about pricing, call get_pricing” — rather than forcing the call unconditionally.

⚠️ Search your whole codebase for tool_choice, not just your main agent file. Retry wrappers, evaluation harnesses and test fixtures are the usual places a forced value survives a migration.

Done when: no request in your codebase sends a forced tool_choice, and your integration tests pass without 400s.

Step 3: Fix Thinking-Block Binding

Fable 5.1 changed how thinking blocks bind to context, and this one breaks quietly rather than loudly.

Each thinking block is valid only against the exact system prompt, tools and conversation history that preceded it. Two consequences:

Editing an earlier turn invalidates every thinking block after it. Any code that rewrites history — summarising old turns, trimming context, injecting a correction — now silently drops reasoning.
Switching models mid-conversation loses blocks the target model cannot read. Fable 5.1 can read thinking blocks produced by earlier Claude models, but a conversation that moves from Fable 5.1 to another model loses them for the turns that run there. Fallback routers and cost-tiering routers are directly affected.

Make the drops visible. Send this beta header:

thinking-binding-controls-2026-08-01

Responses then carry an input_transformations array naming each dropped block with reason: "model_binding_mismatch". Log that array. If it is non-empty in production, your context management is fighting the model.

Done when: your message history is append-only for the turns that carry thinking blocks, and input_transformations is empty on a representative production replay.

Step 4: Replace Assistant Prefill

The third breaking area is assistant prefill — the pattern of seeding the start of the assistant’s reply to force a format. Replace it with structured outputs or a strict tool schema, which is the supported way to constrain shape in Fable 5.1.

Done when: no request seeds assistant content, and output shape is enforced by schema instead.

Step 5: Recalibrate Effort

Reasoning effort does not map one-to-one across versions. A setting tuned on Fable 5 will not produce the same token consumption or latency on 5.1.

Re-run your cost and latency benchmark at each effort level before you trust the old configuration. This is the step teams skip and then discover through an invoice.

Optional but useful: Fable 5.1 supports a thinking display mode that keeps raw reasoning hidden while emitting high-level progress text you can surface to end users:

response = client.beta.messages.create(
    model="claude-fable-5-1",
    thinking={"type": "adaptive", "display": "updates"},
    betas=["thinking-display-updates-2026-08-18"],
)

Previously thinking.display defaulted to omitted, hiding reasoning entirely during multi-tool agent loops. "updates" is a good fit for any UI with a progress indicator.

There is also a mid-conversation-system-clear-at-2026-08-21 beta header for clearing the system prompt mid-conversation — relevant if you swap agent personas inside one thread.

Done when: effort levels are re-benchmarked and your latency/cost budget is confirmed against real traffic.

Step 6: Replay Before You Cut Over

Run matched replays — the same recorded production traffic against both model IDs — and compare output quality, token consumption and error rate. Keep rollback available for at least one full traffic cycle.

Done when: replay results are acceptable and a documented rollback path to claude-fable-5 is tested, not merely assumed.

Migration Checklist

#	Check	Failure mode if skipped
1	Model ID swapped everywhere	Silent use of the old model
2	No forced `tool_choice`	HTTP 400 in production
3	Message history append-only	Thinking blocks silently dropped
4	`input_transformations` logged and empty	Invisible reasoning loss
5	Assistant prefill replaced	Broken output formatting
6	Effort levels re-benchmarked	Cost and latency surprises
7	Matched replay passed	Quality regression at scale
8	Rollback path tested	No way back

Should You Migrate At All?

Not automatically. The upgrade is justified when demanding reasoning or long-horizon work clears a measurable product bar — not because a version number went up. For most traffic, Claude Opus 5 or a cheaper route may be the better quality/latency/cost balance, and Fable 5.1 belongs on the hard subset only.

Decide with a measurement, migrate with the checklist above.

Last verified: September 7, 2026.