Anthropic’s Claude Code AI model encrypts its internal reasoning during extended thinking sessions, returning only a summary instead of the full thought process, according to a June 22 blog post by Patrick McCanna. The model records each session, but the detailed reasoning is stored as a 600-character encrypted signature inaccessible to users without an enterprise agreement.
McCanna inspected the logs of Claude Code’s extended thinking feature and found that the so-called 'thinking blocks' contained only encrypted signatures rather than readable text. Documentation from Anthropic confirms that the API returns a summary of the model’s reasoning, while the full encrypted reasoning is held by Anthropic, which controls the decryption key. Security researcher Matt Green has also analyzed these signature blocks in greater detail.
This approach contrasts with expectations that AI models provide transparent reasoning steps. By encrypting its internal thought process, Claude Code limits user access to the model’s authentic reasoning, raising questions about transparency in AI outputs. The feature is available only under enterprise agreements, restricting broader access to the full reasoning data and potentially affecting trust and verification in AI-assisted decision-making.
Anthropic’s documentation on extended thinking is publicly available at platform.claude.com, and McCanna’s full analysis was published on June 22, 2026, on patrickmccanna.net.