Anthropic said in a research post that new defenses have cut the data distillers harvest from Claude by around 90 to 99%, based on early evidence. Distillation means training a smaller model on a larger model's answers to copy its skills. Anthropic says the most useful data is Claude's written-out reasoning, so its defenses focus there. They include a two-stage filter on its top models, limits on where reasoning can travel, and account bans. Anthropic says it has banned tens of thousands of accounts tied to distillation. It named Zhipu and Moonshot as examples. The Decoder describes both as Chinese companies.
Anthropic measured this by estimating how many reasoning tokens distillers pulled from its own platform, so the figure covers that platform only. The company says it does not know how long the gains will last. It also says it has not yet seen a highly sophisticated, persistent attempt to get around the new defenses. Anthropic expects the contest to keep going back and forth.
For teams building on Claude, the changes have direct effects. Anthropic says its preserved thinking measure, which stops API accounts made on or after August 31, 2026 from editing Claude's earlier context, forced some developers to update their integrations. It also says its filter wrongly blocks fewer than 0.1% of harmless requests. Since September 24, Anthropic has charged for blocked requests in areas where that error rate is low.