Claude Sonnet 4.5 Enforces Conversation Binding Check on Thinking Blocks
Anthropic's Claude Sonnet 4.5 introduces a stricter validation that ties 'thinking blocks' to the exact conversation prefix under which they were generated, triggering a 400 error if the history is modified between requests. Developers running custom agent harnesses may encounter this silently breaking their workflows, as the model verifies a byte-exact match of the prior context before accepting a thinking block. The enforcement is automatic for API accounts created on or after August 31, 2026, across Anthropic, AWS Bedrock, Google Cloud, and Microsoft Foundry; older accounts are currently exempt but recorded. Anthropic cites two reasons for the change: preventing a known model-distillation technique and preserving prompt-cache integrity, since history mutations that break the check also tend to invalidate cached prefixes. Future Claude models are expected to enforce the rule for all accounts regardless of creation date.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in