Claude Code's Visible 'Thinking' Is Just a Summary

The block of text that streams through Claude Code's terminal while the model works through a task — the one labeled 'extended thinking' — is not the model's actual reasoning. Developer Patrick McCanna ran a detailed investigation into Claude Code's session logs and found only a 600-character encrypted signature. No readable thought process. No actual reasoning trace.

What users see when extended thinking is enabled is Anthropic's condensed summary of the model's internal process. The company's documentation confirms it directly: Anthropic encrypts the raw reasoning stream and retains the decryption key. What comes back through the API is a distilled version of that process, not the original. McCanna compares the information loss to converting a high-quality file through a lossy format — something meaningful disappears at each step.

Why this matters if you build with Claude Code

Developers and vibe-coders who use the visible thinking stream to audit agent decisions or debug unexpected outputs are auditing a summary, not the actual logic. If an agent takes a wrong turn mid-task, the displayed 'reasoning' cannot fully account for it — it's a reconstructed account, not a transcript. Anyone who has relied on extended thinking as an audit trail for anything important should factor this in now.

A path to full reasoning does exist: enterprise agreements with Anthropic can unlock unabridged thinking output. For everyone below that tier, transparency here is a UI feature rather than a genuine audit mechanism. This doesn't diminish Claude Code's actual capabilities — the model's performance is driven by its real internal processing, which still produces the final outputs. The limitation is about what you can inspect after the fact. It's a distinction that matters the moment agent reliability becomes your problem.