Anthropic has introduced a new security measure in its Claude Fable 5.1 model to prevent "distillation," a technique where models are trained on the outputs of other models. This new feature, called "preserved thinking," ensures that the conversation history used to generate a thinking block remains consistent. If the history is modified between requests, the model will reject the block, preventing unauthorized training on its outputs. This change primarily affects new accounts using the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, though older accounts can opt-in. AI
IMPACT This change may impact developers building custom agent loops or chat backends that modify conversation history, requiring adjustments to maintain functionality.
RANK_REASON New model release with a novel security feature from a frontier lab. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- Amazon Bedrock
- Anthropic
- claude.ai
- Claude Agent SDK
- Claude API
- Claude Code
- Claude Fable 5.1
- Claude Managed Agents
- Claude Mythos 5.1
- Google Cloud
- Microsoft Foundry
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →