The 'thinking' Claude Code shows you is a summary; the real chain is encrypted
Patrick McCanna went looking for the reasoning behind Claude Code's Extended Thinking and found a summary marketed as transparency. The actual chain, he says, is locked in a ~600-character signature you can't decrypt without Anthropic's key.
McCanna argues the API returns "a SUMMARY of reasoning, NOT the reasoning itself," with the real trace encrypted locally and full thinking gated behind enterprise agreements, so standard users can't audit their agent's decision logic. He likened the lossy step to "saving a bmp as a .jpeg and then editing the .jpeg" (the thread noted he got the lossy direction backwards). Replies pointed out OpenAI documents the same behavior, and that raw traces may not reflect the model's real thoughts anyway.
What you see when Claude "thinks"
McCanna's post makes a blunt claim: the Extended Thinking text Claude Code shows you isn't the model's reasoning, it's a summary of it, presented as if it were transparency. He says the API returns "a SUMMARY of reasoning, NOT the reasoning itself," while the actual chain is compressed into a roughly 600-character signature stored locally that no user can decrypt without Anthropic's key. To make the lossiness vivid he compared it to "saving a bmp as a .jpeg and then editing the .jpeg" — though the thread was quick to note he'd got the direction of the analogy backwards.
The practical complaint underneath the metaphor: full thinking is available only under enterprise agreements, so standard users can't audit how their agent actually reached a decision. He also faults Anthropic's documentation for indirect phrasing that, in his reading, obscures that what you're seeing is only "a summary of Claude's full thinking process."
The thread: norm, billing gripe, or moot point?
The first deflation in the thread was that none of this is secret or unusual.
"no way, the contents of "reasoning_summary" are summarized? fyi openai does the same; not really surprising or particularly evil" — a commenter noting the documented industry norm
A sharper jab quoted Anthropic's own justification back at it.
"Summarized thinking provides the full intelligence benefits of extended thinking, while preventing misuse. > preventing misuse. Imagine not being able to read the tokens you are paying for." — a commenter quoting Anthropic's doc language
The insider rationale came from runeblaze, an applied type theorist working on GenAI, who laid out why hiding the raw chain is deliberate.
"tbh the summarized thinking with encrypted raw thinking is there for many purposes; it is there to: 1. make distillation much harder 2. safety: prevent modifications to the thinking leading to injection attacks. 3. also honestly sometimes the model raw thoughts can be deranged…" — runeblaze, an applied type theorist working on GenAI
And one reply pulled the rug out from under the whole grievance.
"Research shows that even the raw trace tokens do not actually reflect underlying model "thoughts"." — a commenter citing chain-of-thought faithfulness research
If the unfiltered trace was never a faithful record of reasoning, then decrypting it would buy you a more detailed artifact, not a more honest one. The audit McCanna wants may not exist to be gated.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
An Idaho sheriff ran his wife's plate 700 times. The vendor calls stalking rare
Flock says it tracks vehicles, not people. Then a Jerome County, Idaho sheriff queried his own wife's license plate more than 700 times across three months, and the framing fell apart.