The text in Claude Code’s “Extended Thinking” output
patrickmccanna.net
June 22, 2026
1 min read
🔥🔥🔥🔥🔥
63/100
Summary
Claude Code's "Extended Thinking" output includes reasoning logs that are encrypted into a 600-character signature, with Anthropic holding the decryption key. The API provides a summary of the reasoning rather than the full text.
Key Takeaways
- Claude Code encrypts its reasoning into a signature, and Anthropic holds the key, meaning users do not have access to the full reasoning logs.
- The API provides a summary of reasoning rather than the actual reasoning itself, and obtaining full thinking output requires an enterprise agreement.
- The "extended-thinking" output is a summary of the model's thinking logic and does not represent the actual reasoning that influenced the model's actions during a session.
- Users cannot produce a complete record of the reasoning used by their agent during a session, as local reasoning logs are not accessible.
Community Sentiment
Positives
- The discussion highlights a growing awareness of the importance of transparency in AI reasoning, which could lead to better accountability and trust in AI systems.
- Concerns about hidden reasoning in models may drive demand for more explainable AI, ultimately benefiting users who seek clarity in AI decision-making processes.
Concerns
- The lack of transparency in AI reasoning raises significant safety concerns, as hidden processes could allow for prompt injections and other malicious activities.
- Many users express distrust towards models with obfuscated reasoning, fearing that it complicates prompt optimization and increases the risk of misuse.
