Claude Code's 'extended thinking' is a summary, not a true reasoning log

The author warns that the ‘extended thinking’ text Claude Code surfaces via ctrl+o is not the model’s raw chain of reasoning. It is a condensed, after-the-fact summary of the logic Anthropic’s models (Opus/Fable) used during a session, which the author likens to converting a JPEG to BMP, editing it, and passing it off as the original — a lossy process that drops information rather than preserving the real thing.

The practical consequence is about auditability. Anyone who needs a faithful record of why their agent took a given action during a session can’t reconstruct one from local files, because the underlying reasoning traces aren’t exposed to the user. You can scrape a running session’s inputs, outputs, and actions, but that still doesn’t capture the actual reasoning that drove behavior. The author faults Anthropic’s documentation for burying this distinction in soft language — the docs note that extended thinking ‘returns a summary of Claude’s full thinking process,’ which is easy to skim past.

The takeaway for technical readers: don’t promise stakeholders a genuine reasoning audit trail based on what Claude Code displays. The visible thinking is a presentation layer, not forensic evidence of the model’s decision path. The piece closes with a nudge that open-source models need to close the performance gap faster — implying that locally controllable, fully inspectable alternatives matter for anyone with real accountability requirements.