The Illusion of "Extended Thinking" in Claude Code

By patrick | Posted on June 22, 2026 Categories: Mobile Security Mentorship, Inspiring Kids in Tech

The "Extended Thinking" output provided by Claude Code is not authentic.

While the logs appear to contain "thinking blocks"—which are presented as the model's internal reasoning process—a closer inspection reveals a different reality. Upon reviewing these logs over the weekend, I discovered that instead of actual reasoning text, there was merely a signature approximately $L \approx 600$ characters long.

🔍 The Technical Reality

After consulting the official documentation, several critical details emerge:

• Encryption: Claude encrypts its actual reasoning into that specific signature.
• API Limitations: The API does not return the raw reasoning; it provides a SUMMARY of that reasoning.
• Access Control: To obtain the full, unsummarized thinking output, an enterprise agreement is mandatory.

For those interested in a deeper dive, Matt Green has conducted further research and shared more granular observations regarding these signature blocks.

Comparison of Output Types

Feature	"Extended Thinking" (Standard)	Full Thinking (Enterprise)
Content	Summary of logic	Actual raw reasoning
Format	Textual summary	Decrypted logs
Availability	General API / ctrl+o	Enterprise Agreement only
Authenticity	Authentic $\rightarrow$ Derivative	Authentic

⚠️ Critical Warnings

If you are planning to promise a client or stakeholder a complete audit trail, be extremely cautious.

"The extended-thinking output triggered by ctrl+o is merely a summary of the thinking performed by Fable/Opus. It is not the actual cognitive path that drove the model's specific actions during a session."

To put this in perspective, this practice is analogous to:

1. Saving a .jpeg file as a .bmp.
2. Editing the .bmp file.
3. Presenting the result as if it were the original .jpeg.

The Logic Flow

Final Thoughts on Transparency

I find Anthropic's presentation of this behavior to be quite underwhelming. If you require a legitimate record of the logic used by your agent during a session, you cannot produce one using local files.

While you can use "scrappy scraping" to log the following:

• Inputs
• Outputs
• Agent Actions

...even then, you are not capturing the actual reasoning that governed the agent's behavior.

// Example of what you might see in a log vs what you want
{
  "thinking_block": "Encrypted_Signature_600_Chars...", 
  "summary": "I decided to refactor the function to improve readability."
}

If you aren't fully alert (or haven't had your coffee), it is very easy to overlook the fine print stating that "extended thinking returns a summary of Claude’s full thinking process."

Audit Trail Checklist:

• Verify if Enterprise Agreement is active.
• Confirm if logs are raw or summaries.
• Avoid claiming "full transparency" based on local ctrl+o logs.

---

Related Note: Performance gains in Open Source models need to accelerate.

Visuals from the original report: