Editing chat history now breaks Claude’s thinking blocks
Anthropic’s docs give one rule for custom agent loops on Claude Fable 5.1: the messages array must be append-only. Rewriting the system prompt each request, trimming old turns, or editing the tools list invalidates every thinking block after the change, and the API returns a 400 naming the first failing block.
Enforcement is on by default for accounts created on or after August 31, 2026, 00:00 UTC, and Anthropic says later models will enforce it for all users. Setting prefix_mismatch_behavior to “drop_block” with the thinking-binding-controls-2026-08-01 beta header drops the affected blocks instead, unbilled.
Conversations managed by Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK keep the prefix intact.
Up next
90.0% on ARC-AGI-2 at $4.49 per task
ARC Prize’s verified run of Claude Fable 5.1 puts max effort at 97.5% on ARC-AGI-1 for $1.40 per task, while the same setting needs $4.49 per task to reach 90.0% on the harder ARC-AGI-2 set.
Related stories
- Claude now picks its own agent team in Managed Agents
- Claude agents get an OpenTelemetry trail for every step
- Claude Directory opens a submission flow for plugins
- Four Opus 5.5 API changes make Opus 5 requests fail
- Commerce agent blueprint ships with every write staged
- Research preview opens for agent-to-machine control
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.