Anthropic's anti-distillation rule makes agent message history append-only

Anthropic detailed its preserved thinking mechanism, which blocks new API accounts (created on or after August 31, 2026) from editing Claude's prior context. The company says its three-layer defense cut data harvested by distillers by roughly 90–99%. Developers must update integrations: changing the system prompt, tools, or mid-conversation history invalidates all later thinking blocks.

• New API accounts since August 31, 2026 get a 400 error on prefix changes
• Anthropic claims distiller data harvesting dropped by 90–99%
• Two-stage classifier runs on Sonnet 5.5, Opus 5/5.5 and Fable 5/5.1
• Since September 24 Anthropic again charges for requests blocked by safeguards