Anthropic cuts Claude distillation 90–99%

This is less a permanent lock than an escalation in the economics of model copying: Anthropic is making reasoning extraction costly, detectable, and less useful. The trade-off is that legitimate developers must also adapt to tighter controls around thinking blocks.

• –Anthropic says models trained on full reasoning transcripts scored 80.4% on SWE-bench Verified, versus 71.8% when trained on summarized reasoning.
• –Preserved thinking blocks new API accounts from editing prior context, while additional controls restrict reasoning across models, accounts, and organizations.
• –The 90–99% figure is Anthropic’s own platform telemetry, not an independent audit, and the company expects attackers to keep adapting.
• –Tens of thousands of accounts have been banned, while classifiers reportedly maintain a false-positive rate below 0.1%—but blocked requests can still be charged.
• –The broader implication is an industry shift toward identity, behavior, and output-level defenses as model APIs become training-data targets.