Claude Code Extended Thinking: Understanding the Gap Between Reasoning and Summaries

Claude Code's Extended Thinking is a Summary, Not a Raw Log

Claude Code's "Extended Thinking" output is a summarized representation of the model's reasoning logic rather than the actual sequence of thoughts that drove the agent's actions. For standard users, the raw reasoning is encrypted into a signature block (approximately 600 characters long) and stored on the local machine, but the decryption key is held exclusively by Anthropic.

This distinction is critical for developers who rely on these logs for audit trails or debugging. Because the output provided via the ctrl+o command is a summary, it involves data loss—similar to converting a lossless format to a lossy one—meaning the exact logical path the model took is not preserved in the user-facing output.

Access and Limitations of Reasoning Logs

Access to the full, unsummarized thinking output is restricted to users with an enterprise agreement. For all other users, the following limitations apply:

• Local Logs are Inaccessible: The reasoning logs stored on the local system are encrypted and cannot be read by the user.
• API Limitations: The API returns a summary of the reasoning, not the raw reasoning itself.
• Lack of Auditability: It is impossible to produce a complete record of the logic used by a Claude Code agent during a session using local files or standard scraping of inputs and outputs.

Industry Trends in Hidden Reasoning

The practice of hiding raw Chain-of-Thought (CoT) reasoning is not unique to Anthropic; it is a widespread trend among major AI providers, including OpenAI and Google. Industry analysis suggests several primary motivations for this opacity:

Prevention of Model Distillation

A dominant theory among technical observers is that raw reasoning traces are highly valuable for "distillation"—the process where competitors train smaller, cheaper models on the high-quality reasoning chains of frontier models. By encrypting or summarizing these traces, providers protect their R&D investments and maintain a competitive moat.

Brand Protection and "Sane-washing"

Some argue that raw reasoning is often messy, nonsensical, or contains "transient things" that would be alarming or embarrassing to the public if revealed. Summarization allows providers to present a more purposeful and directed version of the model's logic, effectively "sane-washing" the process to avoid admitting that the model may arrive at the correct answer through erratic or inefficient paths.

Technical and Security Implications

Security Risks of Hidden Reasoning

Hidden reasoning introduces specific security vulnerabilities. If a model can perform function calls during a hidden reasoning phase, an attacker could potentially use prompt injection to trigger a secret objective. The reasoning summary could then be used to hide the evidence of this exfiltration from the user, making the attack invisible in the logs.

Debugging and Observability Challenges

The lack of transparency makes it difficult to understand why a model chose a specific, potentially incorrect, approach. In one reported instance, a model decided to create an unnecessarily complex new module because it incorrectly interpreted a project guideline (CLAUDE.md) as a prohibition against modifying existing code. Had the raw thinking process been visible, this error in logic would have been identified immediately, rather than requiring extensive prompting to uncover.

The Nature of LLM "Thinking"

There is an ongoing technical debate regarding whether Chain-of-Thought output actually represents "reasoning" in the human sense. Some suggest that CoT is merely a form of self-reinforcement to encourage stability around beneficial outcomes, meaning that even the raw logs might not provide a true justification for the model's final output.