Anthropic's Fable 5.1 Preserved Thinking locks reasoning context mid-conversation

Anthropic introduced Preserved Thinking in Fable 5.1: thinking blocks are signature-protected, and the API rejects mid-conversation modifications to history (system prompts, tool definitions, or earlier messages) or model switches, wiping all prior reasoning context. Multiple developers discussed the change intensively the same day, arguing its impact is underestimated.

Confirmed

mitsuhiko first disclosed the mechanism: by default it forbids modifying system prompts, tool definitions, or earlier messages mid-conversation, and the API rejects the request on detection.
Switching models mid-conversation is likewise restricted and loses all prior reasoning context (illscience explained the underlying signed-thinking-block mechanism).
The restriction currently applies only to newly onboarded organizations; existing ones are unaffected.
The Fable team (relayed by abenz95 and KyleBoas) says it is still evaluating countermeasures and needs a day or two.

Why it matters

mitsuhiko argued this is fatal for long-running agent sessions: many agent architectures rely on dynamically adjusting or hooking the system prompt, and forbidding history edits makes prompt stability tricky and raises state-persistence problems.
mitsuhiko publicly said he was sad about the change, expecting major impact on developer integrations.

Timeline

09-02: mitsuhiko disclosed the restriction details and flagged risks for long agent sessions; the Fable team responded the same day that it is evaluating options with updates expected in a day or two.

2026-09-02 ~ 2026-09-02 · 6 related posts

Primary sources

Anthropic Locks Reasoning in Fable 5.1: Mid-Conv Edits Now Blocked — mitsuhiko · 2026-09-02 02:15 (UTC+08:00)
Claude Fable 5.1 Preserved Thinking: Preventing Reuse of Reasoning Traces — illscience · 2026-09-02 09:38 (UTC+08:00)
Fable Team Evaluating Anthropic's Inference Restrictions — abenz95 · 2026-09-02 03:12 (UTC+08:00)

[source] Anthropic Locks Reasoning in Fable 5.1: Mid-Conv Edits Now Blocked — mitsuhiko · 2026-09-02
Anthropic Restricts Mid-Conversation Model Switches, Losing Reasoning Context — mitsuhiko · 2026-09-02
Fable 5.1 restricts mid-conversation model switches — KyleBoas_ · 2026-09-02
[source] Fable Team Evaluating Anthropic's Inference Restrictions — abenz95 · 2026-09-02
Anthropic's Model Switch Limitation Sparks Stability Concerns — mitsuhiko · 2026-09-02
[source] Claude Fable 5.1 Preserved Thinking: Preventing Reuse of Reasoning Traces — illscience · 2026-09-02