Visible thinking is a summary of Claude Code.

The long string of characters scrolling by on Claude Code’s terminal screen as the model completes a task (the ‘extended thinking’ output) is not the model’s reasoning. After conducting an exhaustive analysis of Claude Code’s session data, developer Patrick McCanna discovered that all there was a 600 character encrypted hash. Nothing more. There was no actual reasoning, no trace of the actual thought process.

When you turn on ‘extended thinking’, what you’re getting is Anthropic’s summary of what happened inside the model. As their docs make clear, they encrypt the actual reasoning stream and hold onto the decryption key. What you get via the API is a summary of their process, but not the process itself. Think of it like transcoding a high quality file into a lossy format: you lose something with each iteration.

1Why this matters if you build with Claude Code

If you’re using the visible thinking process to audit agents or debug their behavior, you’re auditing a summary, not the logic. When an agent goes off the rails in the middle of a task, you can’t rely on the ‘reasoning’ to explain why. It’s a reconstruction, not a log. So if you’ve been using extended thinking as a log for anything critical, keep this in mind.

But there is a way to get to the bottom of things: if you sign an enterprise agreement with Anthropic, you can get access to unredacted thinking outputs. Until then, transparency is a UI gimmick, not an audit tool. That doesn’t mean Claude Code isn’t powerful — the model’s performance relies on its actual inner workings, which generate the outputs. But you’re limited in what you can see after the fact. And that makes a difference when you care about agent reliability.