Claude Code ‘Extended Thinking’ Output is Summary, Not Authentic Reasoning, Developer Claims
Analysis of session logs suggests visible output suffers from data loss, raising questions about transparency in AI coding agents
A recent blog post by Patrick McCanna, which has gained traction on Hacker News, asserts that the "Extended Thinking" feature in Anthropic’s Claude Code produces a summary of the model's reasoning rather than the authentic, underlying thought process. McCanna argues that while the application records session logs containing raw "thinking blocks," the visible output generated via the `ctrl+o` command is a converted summary that suffers from significant data loss.
McCanna describes the transformation process as akin to saving a JPEG file as a BMP, editing the BMP, and then presenting the result as the original JPEG. He states that the output is a summary of the thinking logic from Fable and Opus models, rather than the actual thinking that drove the model's actions during a session. This distinction is critical for users who may rely on the feature for transparency in automated coding tasks.
The post criticises Anthropic’s documentation for being "indirect," warning users that the feature does not provide a true audit trail of the agent's logic. McCanna notes that the language used in the official docs is subtle enough that users might miss the fact that "extended thinking returns a summary of Claude’s full thinking process" unless they are closely scrutinising the text.
Matt Green is cited in the report for providing further detailed observations on the signature blocks found in the logs. McCanna’s analysis suggests that the session logs contain a signature approximately 600 characters long with no accompanying text, further complicating the ability to trace the model's decision-making path.
The assessment is based on a personal blog post and community comments, rather than an official statement from Anthropic. It remains unclear if this behaviour is consistent across all versions of Claude Code or specific to the models mentioned. The claim that the output is "not authentic" is an interpretation by the blogger, and Anthropic may define "authentic" differently in their technical documentation.
