Anthropic's Oct 12 research post describes three layers against industrial-scale distillation. (1) "Preserved thinking" stops new API accounts (created on or after Aug 31, 2026) from editing prior context around thinking blocks, blocks older models from reading newer models' reasoning, and limits Sonnet 5.5 reasoning to the requesting org. (2) A two-stage classifier, red-teamed in the style of its constitutional classifiers, with a claimed false-positive rate under 0.1%. (3) Stricter limits and bans: tens of thousands of accounts banned, often within two hours. Since Sept 24, blocked requests in low-false-positive categories are billed again. Anthropic says reasoning tokens extracted on its own platform fell 10 to 100+ times. It names Zhipu (GLM-5.3) and Moonshot as distillers and says GLM-5.3's safeguards are easily bypassed. Those claims are Anthropic's own. Anthropic also notes trade-offs: integration changes for developers, rare false blocks, and that it has not yet seen a highly sophisticated attempt to get around the defenses. Defender relevance: API abuse patterns (proxy networks, fraudulent accounts, stolen payment details) look much like other API fraud.