Claude now hides more of its thinking — here's what changed
On Oct 12, 2026, Anthropic rolled out stronger defenses against "distillation" — industrial-scale attempts to copy Claude's reasoning into rival models. The result: Claude summarizes its thinking by default, and requests that try to extract its raw reasoning verbatim now get flagged or blocked.
Anthropic says its new anti-distillation defenses cut the amount of Claude's reasoning that attackers can harvest by 90 to 99%. This matters to you because some ordinary-sounding requests now look suspicious to Claude's safeguards — and knowing which ones will save you a blocked response.
"Distillation" means copying a model's capabilities into a cheaper, smaller model by training it on thousands of that model's responses. The most valuable data for this isn't Claude's final answer — it's Claude's reasoning, the thinking it writes before answering, because that teaches a copycat model how Claude thinks, not just what it says. Attackers run this at scale, often through networks of fake accounts.
Because of this, Claude now summarizes its reasoning before showing it to you, instead of showing the raw thinking. Anthropic tested this: a model trained on full reasoning transcripts scored 80.4% on the SWE-bench Verified coding benchmark; trained on the summarized version, it only scored 71.8%. Summarizing the reasoning makes it far less useful to copy.
Claude also runs a two-stage classifier — the same kind of system used to stop bioweapons or cyberattack requests — to catch attempts to pry out the raw reasoning. It has a false-positive rate under 0.1%, meaning it almost never flags a genuine request by mistake. Since September 24, 2026, Anthropic also charges for requests this classifier blocks. Separately, a feature called "preserved thinking" stops anyone from editing Claude's prior reasoning in a conversation — it applies to API accounts created on or after August 31, 2026, so if your integration used to re-send or tweak an old "thinking" block, that pattern is now rejected.
What this means for you day to day: asking Claude to explain its answer, walk through its steps, or justify a conclusion in its final response is completely normal and still works. What gets blocked is asking Claude to dump its hidden reasoning verbatim, translate old hidden reasoning into another language, or print it inside disguised tags — those are the exact extraction tactics Anthropic named as distillation attempts.
Claude now summarizes its reasoning by default and blocks attempts to extract the raw version verbatim, translated, or disguised — with under a 0.1% false-positive rate. If you want insight into Claude's thinking, ask it to explain its steps in the final answer, not to reveal hidden reasoning. If your API integration edits old "thinking" blocks, check whether "preserved thinking" (accounts from Aug 31, 2026 on) is the reason it's now rejected.