Anthropic says new safeguards have reduced the amount of Claude reasoning its systems identify as extracted by distillers by 10 to 100 times. The company’s account also sets out concrete changes to API access, detection and billing, with some trade-offs for legitimate users.
Anthropic Watch analysis
What happened
In a research post published on 12 October, Anthropic describes three strands of defence: controlling where protected reasoning can be accessed, detecting and blocking extraction attempts, and making repeated attempts harder. It says a two-stage classifier now catches the techniques described in its September report on the company’s most capable generally available models.
Anthropic says the classifier has a false-positive rate below 0.1%, and that reasoning-token extraction on its platform is now 10 to 100 times lower than before its introduction. The company says it has also banned tens of thousands of accounts associated with distillation. These are Anthropic’s measurements and account of its own systems, not an independent evaluation.
Why it matters
The measures affect both would-be extractors and ordinary developers. Anthropic says preserved-thinking changes can require some API integrations to be updated; reasoning transcripts from Claude Sonnet 5.5 are accessible only to the organisation whose request produced them. Since 24 September, the company says it has resumed charging for requests blocked by safeguards in areas where its false-positive rate is below 0.1%.
That makes this more than a security announcement. Model providers are changing how reasoning traces move through products, how suspicious requests are handled and, in some cases, what customers pay for. Anthropic also says its gains may not last: it expects attackers to adapt and acknowledges that its measures can occasionally block benign requests.
Our read
The useful detail is the combination of interventions, not a single impressive-sounding percentage. Anthropic describes changes developers may need to account for and gives a measurable, company-reported result. The unresolved question is how well those protections hold up as attackers change tactics, and whether independent evaluation supports the company’s figures. A chase is not a finish line, however confidently the press release jogs past it.
What to watch
- Whether Anthropic publishes updated results as its safeguards change.
- How often legitimate users encounter blocked requests or integration changes.
- Whether independent testing can assess the reported reduction in extracted reasoning.
Discussion spark: Should providers charge customers for requests their safeguards block, even when the company says the false-positive rate is below 0.1%?
Sources and evidence
- Source update (12 October 2026, 09:30 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.