The Truth Behind Claude Code’s “Extended Thinking” Output

What Happened? News Overview

• Claude Code records each session to disk, capturing the logic of its thought process.
• The outputted “thinking blocks” consist of only a 600-character signature, lacking any actual text.
• To access complete thought output, a corporate contract is required; the API returns only summaries.

Why Is This Important? Key Points to Note

• Claude’s thinking logic is encrypted, with Anthropic holding the keys.
• Users cannot access local files, meaning they cannot verify the actual thought process.
• The output of “Extended Thinking” may mislead users as it does not reflect genuine thought, but rather a summary.

🦈 Shark’s Eye (Curator’s Perspective)

• I believe it’s crucial for users to know that Claude Code’s output is just a summary!
• This means that users seeking accurate records of logic should be wary, as they cannot reconstruct the actual thought process.
• The wording in the documentation is vague, so careful reading is essential!

What’s Next?

• Companies and developers are likely to seek new approaches to address this issue, demanding greater transparency in the thought process.
• There’s a pressing need for improvements in the performance of open-source models, if you ask me!

A Word from Haru-Same

• As a shark reporter, “Haru-Same,” I say that more transparent information sharing is necessary to meet user expectations!

Terminology Explained

• Thinking Block: A part where Claude records its logic, excluding the actual thought process.
• Encryption: A technology that converts data into a specific format to prevent unauthorized access.
• Summary: A concise compilation of information that may lack detailed content.

Source: The text in Claude Code’s “Extended Thinking” output