A blog post by Patrick McCanna took off on Hacker News today, racking up 318 points and over 220 comments at the time of writing. The title says it all: the text you see in Claude Code's Extended Thinking window is not authentic.

What does that mean in practice? When Claude Code is working in "Extended Thinking" mode and you watch it "think" — that scrolling window of reasoning and self-critique — you are not seeing a direct stream from the model's actual process. Anthropic has confirmed that for Claude 4 models, the real thinking is encrypted and stored in a signature field you have no access to. What you see instead is a summary version the model generates after the fact.

You are not seeing what Claude thinks. You are seeing what Claude chooses to tell you it thought.

This isn't exactly new architecture — it has been sitting in Anthropic's documentation for a while. But McCanna puts his finger on something important: people believe they are getting transparency, when what they are actually getting is a presentation layer. That is an enormous difference, especially for developers using this for debugging or trying to understand why the model made a particular decision.

Reactions on HN are split. Some argue that Anthropic is sufficiently clear in its documentation and that this is a problem of user expectations, not one of deception. Others are genuinely frustrated — they have paid for and built workflows around the belief that they were actually seeing "inside" the model. A third camp points out that this is an industry-wide problem: OpenAI, Google, and others do the same thing, but market it just as loosely.

Claude pretends to think — but doesn't show you the real thought - Bilde 1

What makes this particularly interesting right now is the timing. Anthropic has just rolled out the Claude 4 family and its new adaptive thinking system with effort parameters. The more advanced and opaque the backend becomes, the more important it is that what users see actually corresponds to what is happening under the hood.

For the underground perspective: this is not a scandal, but it is an early signal of a growing friction between AI companies' "we are transparent" branding and what transparency actually requires technically. Expect this conversation to resurface every time the next model generation is released.

A note of caution: this is based on community analysis and a single blog post — not independently verified research. But with 318 points on HN, the signal is strong enough to keep an eye on.