Developers who fire up Claude Code expect to watch an AI wrestle with their codebase. They want the stumbles. The false starts. The moments when the model admits a path leads nowhere. What many get instead is something cleaner. Polished. And, according to one sharp-eyed analyst, not the genuine article.
Patrick McCanna laid it out plainly on his blog Sunday. The thinking blocks saved to disk in Claude Code sessions contain no raw reasoning text. Just a long signature string. Roughly 600 characters. Anthropic holds the decryption key. patrickmccanna.net/the-text-in-claude-codes-extended-thinking-output-is-not-authentic/
The local logs don’t deliver the model’s actual chain of thought. They hand back a summary. This distinction matters. Especially for anyone auditing agent behavior or trying to learn from the process.
Anthropic’s architecture separates the thinking from what users see
Official documentation confirms the mechanics. When developers enable extended thinking through the API, they add a thinking object with type enabled and a budget_tokens value. The model produces internal reasoning before the final answer. Yet what returns often isn’t that full internal stream.
“Claude creates thinking content blocks where it outputs its internal reasoning,” the docs state. But they also note that summarized thinking is the default. The visible portion is a condensed version. The full reasoning stays encrypted in the signature field. That signature verifies the blocks came from Claude. Anthropic controls the process. platform.claude.com/docs/en/build-with-claude/extended-thinking
Enterprise customers can access more. Regular users and even many developers on standard plans do not. The summary gets billed based on the original token count. You pay for depth you cannot fully inspect. And in Claude Code specifically, a beta header introduced in February 2026 made redaction the norm.
That header, redact-thinking-2026-02-12, tells the servers to suppress visible reasoning in the terminal and UI. Latency drops. The experience feels snappier. But the reasoning never fully leaves Anthropic’s systems in readable form for the typical session log. A Hacker News discussion and Reddit threads from recent months captured the frustration. Engineers wanted the dead ends. The product delivered an oracle instead.
McCanna didn’t stop at observation. He compared the output to saving a JPEG as a BMP, editing the bitmap, then presenting it as the original image. Data gets lost in translation. The summary reflects the logic. It does not replay the exact sequence that drove the agent’s actions.
But wait. Anthropic announced visible extended thinking back in February 2025 with Claude 3.7 Sonnet. Users could toggle the mode. The company positioned it as a way to build trust. Developers could check the work. Spot contradictions. Understand how the model arrived at answers. anthropic.com/news/visible-extended-thinking
That announcement highlighted benefits for alignment and safety. Visible thoughts could reveal deception. Or at least make it easier to detect. Yet the rollout came with caveats. Faithfulness issues persisted. Models sometimes based decisions on factors not discussed in the visible trace. Safety risks loomed too. Hidden thoughts might enable jailbreaks.
Fast forward to 2026. The product reality in Claude Code tilted toward concealment. A GitHub issue in the claude-code repo from March documented thinking_len=0 after the redact header appeared. Only the encrypted signature remained. Boris from the Claude Code team responded on Hacker News. The change was UI-only, he said. It avoided generating summaries and cut latency. Users could opt out with a settings.json flag.
Even so, the pattern holds. Two major coding tools, Claude Code and competitors using OpenAI’s reasoning models, quietly shifted toward hiding the full process. A detailed Reddit post from May outlined the trade-offs. Developers pay for reasoning tokens yet receive redacted or summarized output. Alternatives like Aider or Cline with local models suddenly looked more attractive precisely because they showed every step.
The implications stretch beyond individual frustration. Audit trails matter in regulated industries. If an AI agent makes a critical decision in financial code or medical systems, the record should reflect actual reasoning. Summaries fall short. So do encrypted blobs only decryptable by the vendor.
Matt Green, a cryptography researcher, examined the signature blocks in more depth. His analysis, linked by McCanna, reinforces that the design keeps the real reasoning under Anthropic’s control. Your machine sees the receipt. Not the full ledger.
And the language in the documentation? McCanna called it indirect. A reader skimming after too little coffee could miss the crucial qualifier. The extended thinking output from certain interfaces is a summary of the full process. Not the process itself.
This isn’t deception in the classic sense. The model still performs the computation. Performance gains appear real. Agents handle complex tasks better with the extra tokens allocated to thought. Yet the transparency once promised has narrowed.
Open source models continue to close the gap. Their reasoning stays fully visible by design. No vendor gatekeeping the logs. No paid tiers unlocking what should be fundamental for serious users. McCanna ended his post with a call for faster progress there. Many developers appear to agree.
Recent discussions on X amplified the original post. Hacker News threads from today drove traffic to McCanna’s analysis. The timing feels pointed. As coding agents move from novelty to daily infrastructure, the rules around visibility deserve scrutiny.
Enterprise agreements unlock fuller access. That fact itself reveals the tiered nature of the offering. Teams with deep pockets and legal leverage see more. Everyone else works with the summary version. The one that looks authentic enough in a UI but won’t stand up to forensic inspection.
So the next time a Claude Code session logs its thoughts, remember what those blocks actually contain. A verified, encrypted placeholder. A condensed narrative. Proof that reasoning happened. But not the reasoning itself. Developers chasing true understanding may need to look elsewhere. Or demand better defaults from the platforms shaping their tools.
WebProNews is an iEntry Publication