Anthropic Says New Defenses Cut Claude Distillation by Up to 99%

Anthropic says early evidence suggests its strengthened safeguards have reduced the amount of data harvested by distillers by around 90 to 99%, according to a company research post. Anthropic estimates how many tokens of Claude's reasoning distillers extract on its own platform, and says that number is now consistently 10 to 100 times lower than before its new classifier.

The defenses work in three ways. Anthropic restricts where reasoning can go: with Claude Fable 5.1 and Claude Opus 5.5 it introduced preserved thinking, which stops new API accounts from editing Claude's prior context. It runs a two stage classifier on its most capable generally available models that it says catches all the techniques described in its September report on distillation campaigns, with a false positive rate below 0.1%. And it applies stricter usage limits to accounts that keep trying to extract reasoning, banning tens of thousands of accounts, often within two hours of a campaign starting.

Anthropic says the measures involve trade offs, including turning away meaningful revenue and trillions of tokens of paying traffic from distillers. It expects attackers to find ways around the safeguards and says it updates them quickly when it observes a new technique.