Claude Fable 5.1 Has Four Breaking Changes, Not Three

Anthropic shipped Claude Fable 5.1 on Sept. 1 with four breaking API changes. The migration guides written since have been counting three.

The fourth is a data retention requirement, and it is the one most likely to stop a deployment rather than break a request.

The change nobody counted

Anthropic's API release notes state that Claude Fable 5.1 and Claude Mythos 5.1 "require 30-day data retention" and are "not available under zero data retention unless expressly authorized by Anthropic."

That is not a parameter to fix in a client library. Zero data retention is the configuration organisations adopt when a contract, a regulator or an internal policy forbids the vendor from holding request content, and it is common in health, finance and public sector work. On the two newest models, it now needs written permission from Anthropic.

Every migration write-up in the top search results for the release counts three breaking changes and lists the same three. None of them mention retention.

What the other three break

The three that are being counted are real, and they hit different kinds of code.

Forced tool use is gone. On claude-fable-5-1, a tool_choice of {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 invalid_request_error. Only auto and none remain. Anthropic points callers at strict tool use or structured outputs to get schema-conformant arguments instead.

Thinking blocks are now bound to the model that produced them. Fable 5.1 will read blocks from Claude Opus 5, Claude Fable 5, Claude Mythos 5 and earlier models, but those earlier models cannot read Fable 5.1's, and the API drops blocks replayed to them.

Sampling parameters are refused. Setting temperature, top_p or top_k to anything other than the default returns a 400, matching the behaviour Anthropic introduced on Claude Opus 4.8.

The cutoff that decides whether you notice

The thinking-block change carries a date test rather than a version test. For accounts created on or after Aug. 31, 2026, replaying a thinking block after the system prompt, the tools or an earlier message has changed returns a 400. Accounts created before that date do not get the error.

Two teams on the same model, running the same code, will therefore see different behaviour depending on when the account was opened. A beta header, thinking-binding-controls-2026-08-01, exposes an input_transformations field reporting which blocks were dropped.

Anything that edits conversation history and replays it is exposed: agent runtimes that resume sessions, tool routers, and anything that trims context before a retry.

Cheaper reads, unchanged list price

Fable 5.1 lists at $10 per million input tokens and $50 per million output tokens, the same as Fable 5. The context window is 1 million tokens by default with 128,000 maximum output tokens.

The price that moved is caching. Cache reads on Fable 5.1 and Mythos 5.1 are $0.25 per million tokens, which Anthropic puts at 0.025 times the base input price against 0.1 times on its other models. Cache writes are unchanged.

For a workload that re-reads a large cached prefix on every call, that is a four-fold cut on the dominant line of the bill, and it arrived without a change to the headline rate that most comparisons quote.

Mythos 5.1 ships as claude-mythos-5-1 to participants in a programme Anthropic calls Project Glasswing, with the same context window, output limit and pricing.

Anthropic has not published a date for when the retention requirement might be relaxed, and the release notes give no route to an exception beyond asking.

Related: our reporting on vendor-supplied benchmark scores in the GPT-6 Astra release and on the CVE totals nobody can reconcile.